Merge branch 'main' of https://code.sdsdev.co.kr/dx-tan/SBT-IDP into optimize_performance

This commit is contained in:
Viet Anh Nguyen 2023-12-14 13:40:12 +07:00
commit 6873ffce05
357 changed files with 2582287 additions and 13195 deletions

6
.gitignore vendored
View File

@ -12,6 +12,11 @@ backup/
*.log
__pycache__
# migrations/
sdsvtd copy/
oickle
*.sample
._bit
__pycache__
test/
._git/
sdsvkvu_/
@ -21,3 +26,4 @@ postgres_data/
curl.md
cope2n-api/fwd_api/commands/init_database.py
/data
backup

0
cope2n-ai-fi/._gitmodules Normal file → Executable file
View File

0
cope2n-ai-fi/Dockerfile_fwd Normal file → Executable file
View File

0
cope2n-ai-fi/TODO.md Normal file → Executable file
View File

0
cope2n-ai-fi/api/manulife/predict_manulife.py Normal file → Executable file
View File

View File

@ -1,4 +1,5 @@
import cv2
import nltk
import urllib
import random
import numpy as np
@ -7,6 +8,8 @@ import uuid
import sys, os
cur_dir = str(Path(__file__).parents[2])
sys.path.append(cur_dir)
nltk.data.path.append(os.path.join((os.getcwd() + '/nltk_data')))
from modules.sdsvkvu import load_engine, process_img
from modules.ocr_engine import OcrEngine
from configs.sdsap_sbt import device, ocr_cfg, kvu_cfg
@ -22,8 +25,10 @@ def load_ocr_engine(opt) -> OcrEngine:
print("OCR engine configfs: \n", ocr_cfg)
print("KVU configfs: \n", kvu_cfg)
ocr_engine = load_ocr_engine(ocr_cfg)
kvu_cfg['ocr_engine'] = ocr_engine
# ocr_engine = load_ocr_engine(ocr_cfg)
# kvu_cfg['ocr_engine'] = ocr_engine
kvu_cfg['ocr_configs'] = ocr_cfg
option = kvu_cfg['option']
kvu_cfg.pop("option") # pop option
sbt_engine = load_engine(kvu_cfg)

0
cope2n-ai-fi/common/AnyKey_Value/utils/split_docs.py Normal file → Executable file
View File

0
cope2n-ai-fi/configs/default_env.py Normal file → Executable file
View File

0
cope2n-ai-fi/configs/manulife/__init__.py Normal file → Executable file
View File

0
cope2n-ai-fi/configs/manulife/configs.py Normal file → Executable file
View File

0
cope2n-ai-fi/configs/sdsap_sbt/__init__.py Normal file → Executable file
View File

0
cope2n-ai-fi/configs/sdsap_sbt/configs.py Normal file → Executable file
View File

0
cope2n-ai-fi/modules/__init__.py Normal file → Executable file
View File

View File

@ -1,24 +0,0 @@
# Model weights
weights/
microsoft/
nltk_data/
# Visualize
visualize
# External
sdsvkvu/externals/ocr_engine_deskew/externals/
#
__pycache__
*/__pycache__
*/*/__pycache__
#
.git_temp/
# Packages
build/
dist/

View File

@ -1,4 +0,0 @@
[submodule "sdsvkvu/externals/basic_ocr"]
path = sdsvkvu/externals/basic_ocr
url = https://code.sdsdev.co.kr/tuanlv/IDP-BasicOCR.git

View File

@ -1,13 +0,0 @@
Copyright 2023 tuanlv
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.

View File

@ -1,2 +0,0 @@
include sdsvkvu/weights/*/*.yaml
include sdsvkvu/weights/*/checkpoints/best_model.pth

View File

@ -1,122 +0,0 @@
<p align="center">
<h1 align="center">SDSVKVU</h1>
</p>
***Feature***
- Extract pairs of key-value in documents: Invoice/Receipt, Forms, Government documents (Id cards, driver license, birth's certificate)
- Language: VI + EN
***What's news***
### - Ver 0.0.1:
- Support inputs: image, PDF file (single or multi pages)
- Extract all pairs key-value return raw_outputs
+ Weights: weights/key_value_understanding-20230716-085549_final
- For VAT invoices : Extract 14 specific fields
+ Weights: weights/key_value_understanding-20230627-164536_fi
- For SBT invoices ("sbt" option): Extract table in SBT invoice
+ Weights: weights/key_value_understanding-20230812-170826_sbt_2
### - Ver 0.0.2: Add more option: "vtb" - Vietin Bank
- For Vietin Bank document ("vtb" option): Extract 6 specific fileds
+ Weights: weights/key_value_understanding-20230824-164236_vietin
### - Ver 0.0.3: Add default option:
- Return all potential pairs of key-value, title, only key, triplet, and table with raw key
### - Ver 0.0.4: Add more option: "manulife" - Manulife Issurance
- For Manulife Insurance document ("manulife" option): Extract all potential pairs of key-value, title, only key, triplet, and table with raw key + Type of medical documents
+ Weights: weights/key_value_understanding-20231024-125646_manulife2
### Ver 0.1.0: Modify KVU model for SBT
### - Ver 0.1.0: Add option: "sbt_v2" - SBT project
- For SBT imei/invoice ("sbt_v2" option): Extract 4 specific fields
+ Weights: weights/key_value_understanding_for_sbt-20231108-143935
## I. Setup
***Dependencies***
- Python: 3.10
- Torch: 1.11.3
- CUDA: 11.6
- transformers: 4.30.0
```
pip install -v -e .
```
## II. Inference
run cmd: python test.py
```
import os
from sdsvkvu import load_engine, process_img
os.environ["CUDA_VISIBLE_DEVICES"]="1"
if __name__ == "__main__":
kwargs = {"device": "cuda:0"}
img_dir = "/mnt/ssd1T/tuanlv/02-KVU/sdsvkvu/visualize/test_img/RedInvoice_WaterPurfier_Feb_PVI_829_0.jpg"
save_dir = "/mnt/ssd1T/tuanlv/02-KVU/sdsvkvu/visualize/test2/"
engine = load_engine(kwargs)
# option: "vat" for vat invoice outputs, "sbt": sbt invoice outputs, else for raw outputs
outputs = process_img(img_dir, save_dir, engine, export_all=False, option="vat")
```
# Structure project
.
├── sdsvkvu
│   ├── main.py
├── externals
│   │   ├── __init__.py
│   │   ├── basic_ocr
│   │   │   ├── ...
│   │   ├── ocr_engine
│   │   │   ├── ...
│   │   ├── ocr_engine_deskew
│   │   │   ├── ...
│   ├── model
│   │   ├── combined_model.py
│   │   ├── document_kvu_model.py
│   │   ├── __init__.py
│   │   ├── kvu_model.py
│   │   └── relation_extractor.py
│   ├── modules
│   │   ├── __init__.py
│   │   ├── predictor.py
│   │   ├── preprocess.py
│   │   └── run_ocr.py
│   ├── requirements.txt
│   ├── settings.yml
│   ├── sources
│   │   ├── __init__.py
│   │   ├── kvu.py
│   │   └── utils.py
│   ├── utils
│   │   ├── dictionary
│   │   │   ├── __init__.py
│   │   │   ├── sbt.py
│   │   │   └── vat.py
│   │   │   └── vtb.py
│   │   │   ├── manulife.py
│   │   │   ├── sbt_v2.py
│   │   ├── __init__.py
│   │   ├── post_processing.py
│   │   ├── query
│   │   │   ├── __init__.py
│   │   │   ├── sbt.py
│   │   │   └── vat.py
│   │   │   └── vtb.py
│   │   │   ├── all.py
│   │   │   ├── manulife.py
│   │   │   ├── sbt_v2.py
│   │   └── utils.py
├── weights
│   └── key_value_understanding-20230627-164536_fi
│   ├── key_value_understanding-20230812-170826_sbt_2
│   └── key_value_understanding-20230716-085549_final
│   └── key_value_understanding-20230824-164236_vietin
│   └── key_value_understanding-20231024-125646_manulife2
│   └── key_value_understanding_for_sbt-20231108-143935
├── LICENSE
├── MANIFEST.in
├── pyproject.toml
├── README.md
├── scripts
│   └── run.sh
├── setup.cfg
├── setup.py
├── test.py
└── visualize

View File

@ -1,8 +0,0 @@
import os
import sys
from pathlib import Path
cur_dir = str(Path(__file__).parents[0])
sys.path.append(cur_dir)
sys.path.append(os.path.join(cur_dir, "sdsvkvu"))
from sdsvkvu import load_engine, process_img, process_pdf, process_dir

Binary file not shown.

Before

Width:  |  Height:  |  Size: 396 KiB

View File

@ -1,6 +0,0 @@
[build-system]
requires = [
"setuptools>=65",
"wheel"
]
build-backend = "setuptools.build_meta"

View File

@ -1,28 +0,0 @@
nltk
six
deskew
jdeskew
pdf2image
omegaconf
imagesize
xmltodict
dicttoxml
terminaltables
Pillow>=9.4.0
nptyping==1.4.2
opencv-python==4.5.4.60 ##
opencv-python-headless==4.5.4.60
overrides==4.1.2
# transformers==4.30.0
sentencepiece==0.1.99
seqeval==0.0.12
tensorboard>=2.2.0
scipy==1.9.1
# code-style
isort==5.9.3
black==21.9b0
# pytorch
# --find-links https://download.pytorch.org/whl/torch_stable.html
# torch==1.13.1+cu116
# torchvision==0.14.1+cu116
tldextract==5.1.1

View File

@ -1,26 +0,0 @@
cd /mnt/hdd4T/OCR/tuanlv/02-KVU/sdsvkvu
export CUDA_VISIBLE_DEVICES=0
python sdsvkvu/main.py \
--img_dir /mnt/hdd4T/OCR/tuanlv/00-Datasets/SBT_DATA/invoice_validation \
--save_dir /mnt/hdd4T/OCR/tuanlv/02-KVU/02-KVU_test/visualize/sbt_invoice \
--kvu_params "{\"device\":\"cuda:0\"}" \
--doc_type "sbt_v2" \
--export_img 1
# python sdsvkvu/main.py \
# --img_dir /mnt/hdd4T/OCR/tuanlv/00-Datasets/SBT_DATA/imei_validation \
# --save_dir /mnt/hdd4T/OCR/tuanlv/02-KVU/sdsvkvu/visualize/test_sbt_imei \
# --kvu_params "{\"device\":\"cuda:0\"}" \
# --doc_type "sbt_v2" \
# --export_img 1
# python sdsvkvu/main.py \
# --img_dir /mnt/hdd4T/OCR/tuanlv/02-KVU/sdsvkvu/visualize/test_sbt2 \
# --save_dir /mnt/hdd4T/OCR/tuanlv/02-KVU/sdsvkvu/visualize/test_sbt2 \
# --kvu_params "{\"device\":\"cuda:0\"}" \
# --doc_type "sbt_v2" \
# --export_img 1

View File

@ -1,122 +0,0 @@
Metadata-Version: 2.1
Name: sdsvkvu
Version: 0.0.1
Summary: SDSV OCR Team: Key-value understanding
Home-page: https://github.com/open-mmlab/mmocr
Author: tuanlv
Author-email: lv.tuan3@samsung.com
License: Apache License 2.0
Classifier: Development Status :: 4 - Beta
Classifier: License :: OSI Approved :: Apache Software License
Classifier: Operating System :: OS Independent
Classifier: Programming Language :: Python :: 3.9
Requires-Python: >=3.9
Description-Content-Type: text/markdown
License-File: LICENSE
<p align="center">
<h1 align="center">SDSVKVU</h1>
</p>
***Feature***
- Extract pairs of key-value in documents: Invoice/Receipt, Forms, Government documents (Id cards, driver license, birth's certificate)
- Language: VI + EN
***What's news***
### - Ver 0.0.1:
- Support inputs: image, PDF file (single or multi pages)
- Extract all pairs key-value return raw_outputs
+ Weights: sdsvkvu/weights/key_value_understanding-20230716-085549_final
- For VAT invoices : Extract 14 specific fields
+ Weights: sdsvkvu/weights/key_value_understanding-20230627-164536_fi
- For SBT invoices ("sbt" option): Extract table in SBT invoice
+ Weights: sdsvkvu/weights/key_value_understanding-20230617-162324_sbt
### - Ver 0.0.2: Add more option: "vtb" - Vietin Bank
- For Vietin Bank document ("vtb" option): Extract 6 specific fileds
+ Weights: sdsvkvu/weights/key_value_understanding-20230824-164236_vietin
### - Ver 0.0.3: Add default option:
- Return all potential pairs of key-value, title, only key, triplet, and table with raw key
## I. Setup
***Dependencies***
- Python: 3.10
- Torch: 1.11.3
- CUDA: 11.6
- transformers: 4.30.0
```
pip install -v -e .
```
## II. Inference
run cmd: python test.py
```
import os
from sdsvkvu import load_engine, process_img
os.environ["CUDA_VISIBLE_DEVICES"]="1"
if __name__ == "__main__":
kwargs = {"device": "cuda:0"}
img_dir = "/mnt/ssd1T/tuanlv/02-KVU/sdsvkvu/visualize/test_img/RedInvoice_WaterPurfier_Feb_PVI_829_0.jpg"
save_dir = "/mnt/ssd1T/tuanlv/02-KVU/sdsvkvu/visualize/test2/"
engine = load_engine(kwargs)
# option: "vat" for vat invoice outputs, "sbt": sbt invoice outputs, else for raw outputs
outputs = process_img(img_dir, save_dir, engine, export_all=False, option="vat")
```
# Structure project
.
├── sdsvkvu
│   ├── main.py
├── externals
│   │   ├── __init__.py
│   │   ├── ocr_engine
│   │   │   ├── ...
│   │   ├── ocr_engine_deskew
│   │   │   ├── ...
│   ├── model
│   │   ├── combined_model.py
│   │   ├── document_kvu_model.py
│   │   ├── __init__.py
│   │   ├── kvu_model.py
│   │   └── relation_extractor.py
│   ├── modules
│   │   ├── __init__.py
│   │   ├── predictor.py
│   │   ├── preprocess.py
│   │   └── run_ocr.py
│   ├── requirements.txt
│   ├── settings.yml
│   ├── sources
│   │   ├── __init__.py
│   │   ├── kvu.py
│   │   └── utils.py
│   ├── utils
│   │   ├── dictionary
│   │   │   ├── __init__.py
│   │   │   ├── sbt.py
│   │   │   └── vat.py
│   │   │   └── vtb.py
│   │   ├── __init__.py
│   │   ├── post_processing.py
│   │   ├── query
│   │   │   ├── __init__.py
│   │   │   ├── sbt.py
│   │   │   └── vat.py
│   │   │   └── vtb.py
│   │   └── utils.py
│   └── weights
│   └── key_value_understanding-20230627-164536_fi
│   ├── key_value_understanding-20230617-162324_sbt
│   └── key_value_understanding-20230716-085549_final
│   └── key_value_understanding-20230824-164236_vietin
├── LICENSE
├── MANIFEST.in
├── pyproject.toml
├── README.md
├── scripts
│   └── run.sh
├── setup.cfg
├── setup.py
├── test.py
└── visualize

View File

@ -1,49 +0,0 @@
LICENSE
MANIFEST.in
README.md
pyproject.toml
setup.cfg
setup.py
sdsvkvu/__init__.py
sdsvkvu/main.py
sdsvkvu.egg-info/PKG-INFO
sdsvkvu.egg-info/SOURCES.txt
sdsvkvu.egg-info/dependency_links.txt
sdsvkvu.egg-info/not-zip-safe
sdsvkvu.egg-info/requires.txt
sdsvkvu.egg-info/top_level.txt
sdsvkvu/externals/__init__.py
sdsvkvu/externals/basic_ocr/__init__.py
sdsvkvu/externals/basic_ocr/run.py
sdsvkvu/externals/ocr_engine/__init__.py
sdsvkvu/externals/ocr_engine/run.py
sdsvkvu/externals/ocr_engine_deskew/__init__.py
sdsvkvu/externals/ocr_engine_deskew/run.py
sdsvkvu/model/__init__.py
sdsvkvu/model/combined_model.py
sdsvkvu/model/document_kvu_model.py
sdsvkvu/model/kvu_model.py
sdsvkvu/model/relation_extractor.py
sdsvkvu/model/sbt_model.py
sdsvkvu/modules/__init__.py
sdsvkvu/modules/predictor.py
sdsvkvu/modules/preprocess.py
sdsvkvu/modules/run_ocr.py
sdsvkvu/sources/__init__.py
sdsvkvu/sources/kvu.py
sdsvkvu/sources/utils.py
sdsvkvu/utils/__init__.py
sdsvkvu/utils/post_processing.py
sdsvkvu/utils/utils.py
sdsvkvu/utils/word2line.py
sdsvkvu/utils/dictionary/__init__.py
sdsvkvu/utils/dictionary/sbt.py
sdsvkvu/utils/dictionary/sbt_v2.py
sdsvkvu/utils/dictionary/vat.py
sdsvkvu/utils/dictionary/vtb.py
sdsvkvu/utils/query/__init__.py
sdsvkvu/utils/query/all.py
sdsvkvu/utils/query/sbt.py
sdsvkvu/utils/query/sbt_v2.py
sdsvkvu/utils/query/vat.py
sdsvkvu/utils/query/vtb.py

View File

@ -1,21 +0,0 @@
nltk
six
deskew
jdeskew
pdf2image
omegaconf
imagesize
xmltodict
dicttoxml
terminaltables
Pillow==9.4.0
nptyping==1.4.2
opencv-python==4.5.4.60
opencv-python-headless==4.5.4.60
overrides==4.1.2
sentencepiece==0.1.99
seqeval==0.0.12
tensorboard>=2.2.0
scipy==1.9.1
isort==5.9.3
black==21.9b0

View File

@ -1,4 +0,0 @@
from .main import load_engine
from .main import process_img
from .main import process_pdf
from .main import process_dir

View File

@ -1,6 +0,0 @@
__pycache__/
visualize/
results/
*.jpeg
*.jpg
*.png

View File

@ -1,47 +0,0 @@
# OCR Engine
OCR Engine is a Python package that combines text detection and recognition models from [mmdet](https://github.com/open-mmlab/mmdetection) and [mmocr](https://github.com/open-mmlab/mmocr) to perform Optical Character Recognition (OCR) on various inputs. The package currently supports three types of input: a single image, a recursive directory, or a csv file.
## Installation
To install OCR Engine, clone the repository and install the required packages:
```bash
git clone git@github.com:mrlasdt/ocr-engine.git
cd ocr-engine
pip install -r requirements.txt
```
## Usage
To use OCR Engine, simply run the `ocr_engine.py` script with the desired input type and input path. For example, to perform OCR on a single image:
```css
python ocr_engine.py --input_type image --input_path /path/to/image.jpg
```
To perform OCR on a recursive directory:
```css
python ocr_engine.py --input_type directory --input_path /path/to/directory/
```
To perform OCR on a csv file:
```
python ocr_engine.py --input_type csv --input_path /path/to/file.csv
```
OCR Engine will automatically detect and recognize text in the input and output the results in a CSV file named `ocr_results.csv`.
## Contributing
If you would like to contribute to OCR Engine, please fork the repository and submit a pull request. We welcome contributions of all types, including bug fixes, new features, and documentation improvements.
## License
OCR Engine is released under the [MIT License](https://opensource.org/licenses/MIT). See the LICENSE file for more information.

View File

@ -1,10 +0,0 @@
☐ refactor argument parser of run.py
☐ add timer level, logging level and write_mode to argumments
☐ add paddleocr deskew to the code
☐ fix the deskew code to resize the image only for detecting the angle, we want to feed the original size image to the text detection pipeline so that the bounding boxes would be mapped back to the original size
☐ ocr engine import took too long
☐ add word level to write_mode
☐ add word group and line
change max_x_dist from pixel to percentage of box width
☐ visualization: adjust fontsize dynamically

View File

@ -1,11 +0,0 @@
# # Define package-level variables
# __version__ = '0.0'
# Import modules
from .src.ocr import OcrEngine
# from .src.word_formation import words_to_lines
from .src.word_formation import words_to_lines_tesseract as words_to_lines
from .src.utils import ImageReader, read_ocr_result_from_txt
from .src.dto import Word, Line, Page, Document, Box
# Expose package contents
__all__ = ["OcrEngine", "Box", "Word", "Line", "Page", "Document", "words_to_lines", "ImageReader", "read_ocr_result_from_txt"]

View File

@ -1,9 +0,0 @@
output*
*.pyc
*.jpg
check
weights/
workdirs/
__pycache__*
test_hungbnt.py
libs*

View File

@ -1,29 +0,0 @@
<p align="center">
<h1 align="center">Dewarp</h1>
</p>
***Feature***
- Align document
## I. Setup
***Dependencies***
- Python: 3.8
- Torch: 1.10.2
- CUDA: 11.6
- transformers: 4.28.1
### 1. Install PaddlePaddle
```
python -m pip install paddlepaddle-gpu==2.4.2.post116 -f https://www.paddlepaddle.org.cn/whl/linux/mkl/avx/stable.html
```
### 2. Install sdsv_dewarp
```
pip install -v -e .
```
## II. Test
```
python test.py --input samples --out demo/outputs --device 'cuda'
```

View File

@ -1,3 +0,0 @@
model_dir: /mnt/hdd4T/OCR/tuanlv/01-BasicOCR/ocr-engine-deskew/externals/sdsv_dewarp/weights/ch_ppocr_mobile_v2.0_cls_infer
gpu_mem: 3000
max_batch_size: 32

View File

@ -1,8 +0,0 @@
model_dir: /mnt/hdd4T/OCR/tuanlv/01-BasicOCR/ocr-engine-deskew/externals/sdsv_dewarp/weights/ch_PP-OCRv3_det_infer
gpu_mem: 3000
det_limit_side_len: 1560
det_limit_type: max
det_db_unclip_ratio: 1.85
det_db_thresh: 0.3
det_db_box_thresh: 0.5
det_db_score_mode: fast

View File

@ -1,7 +0,0 @@
paddleocr>=2.0.1
opencv-contrib-python
opencv-python
numpy
gdown==3.13.0
imgaug==0.4.0

View File

@ -1,45 +0,0 @@
Metadata-Version: 2.1
Name: sdsv-dewarp
Version: 1.0.0
Summary: Dewarp document
Home-page:
License: Apache License 2.0
Classifier: Development Status :: 4 - Beta
Classifier: License :: OSI Approved :: Apache Software License
Classifier: Operating System :: OS Independent
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.6
Classifier: Programming Language :: Python :: 3.7
Classifier: Programming Language :: Python :: 3.8
Classifier: Programming Language :: Python :: 3.9
Description-Content-Type: text/markdown
<p align="center">
<h1 align="center">Dewarp</h1>
</p>
***Feature***
- Align document
## I. Setup
***Dependencies***
- Python: 3.8
- Torch: 1.10.2
- CUDA: 11.6
- transformers: 4.28.1
### 1. Install PaddlePaddle
```
python -m pip install paddlepaddle-gpu==2.4.2.post116 -f https://www.paddlepaddle.org.cn/whl/linux/mkl/avx/stable.html
```
### 2. Install sdsv_dewarp
```
pip install -v -e .
```
## II. Test
```
python test.py --input samples --out demo/outputs --device 'cuda'
```

View File

@ -1,15 +0,0 @@
README.md
setup.py
sdsv_dewarp/__init__.py
sdsv_dewarp/api.py
sdsv_dewarp/config.py
sdsv_dewarp/factory.py
sdsv_dewarp/models.py
sdsv_dewarp/utils.py
sdsv_dewarp/version.py
sdsv_dewarp.egg-info/PKG-INFO
sdsv_dewarp.egg-info/SOURCES.txt
sdsv_dewarp.egg-info/dependency_links.txt
sdsv_dewarp.egg-info/not-zip-safe
sdsv_dewarp.egg-info/requires.txt
sdsv_dewarp.egg-info/top_level.txt

View File

@ -1,6 +0,0 @@
paddleocr>=2.0.1
opencv-contrib-python
opencv-python
numpy
gdown==3.13.0
imgaug==0.4.0

View File

@ -1,200 +0,0 @@
import math
import numpy as np
from typing import List
import cv2
import collections
import logging
import imgaug.augmenters as iaa
from imgaug.augmentables.polys import Polygon, PolygonsOnImage
from sdsv_dewarp.models import PaddleTextClassifier, PaddleTextDetector
from sdsv_dewarp.config import Cfg
from .utils import *
MIN_LONG_EDGE = 40**2
NUMBER_BOX_FOR_ALIGNMENT = 200
MAX_ANGLE = 180
MIN_ANGLE = 1
MIN_NUM_BOX_TEXT = 3
CROP_SIZE = 3000
logging.basicConfig(level=logging.INFO)
LOGGER = logging.getLogger(__name__)
class AlignImage:
"""Rotate image to 0 degree
Args:
text_detector (deepmodel): Text detection model
text_cls (deepmodel): Text classification model (0 or 180)
Return:
is_blank (bool): Blank image when haven't boxes text
image_align: Image after alignment
angle_align: Degree of angle alignment
"""
def __init__(self, text_detector: dict, text_cls: dict, device: str = 'cpu'):
self.text_detector = None
self.text_cls = None
self.use_gpu = True if device != 'cpu' else False
self._init_model(text_detector, text_cls)
def _init_model(self, text_detector, text_cls):
det_config = Cfg.load_config_from_file(text_detector['config'])
det_config['model_dir'] = text_detector['weight']
cls_config = Cfg.load_config_from_file(text_cls['config'])
cls_config['model_dir'] = text_cls['weight']
self.text_detector = PaddleTextDetector(config=det_config, use_gpu=self.use_gpu)
self.text_cls = PaddleTextClassifier(config=cls_config, use_gpu=self.use_gpu)
def _cal_width(self, poly_box):
"""Calculate width of a polygon [[x1, y1], [x2, y2], [x3, y3], [x4, y4]]"""
tl, tr, br, bl = poly_box
edge_s, edge_l = distance(tl, tr), distance(tr, br)
return max(edge_s, edge_l)
def _get_most_frequent(self, values):
values = np.array(values)
# create the histogram
hist, bins = np.histogram(values, bins=np.arange(0, 181, 10))
# get the index of the most frequent angle
index = np.argmax(hist)
# get the most frequent angle
most_frequent_angle = (bins[index] + bins[index + 1]) / 2
return most_frequent_angle
def _cal_angle(self, poly_box):
"""Calculate the angle between two point"""
a = poly_box[0]
b = poly_box[1]
c = poly_box[2]
# Get the longer edge
if distance(a, b) >= distance(b, c):
x, y = a, b
else:
x, y = b, c
angle = math.degrees(math.atan2(-(y[1] - x[1]), y[0] - x[0]))
if angle < 0:
angle = 180 - abs(angle)
return angle
def _reject_outliers(self, data, m=5.0):
"""Remove noise angle"""
list_index = np.arange(len(data))
d = np.abs(data - np.median(data))
mdev = np.median(d)
s = d / (mdev if mdev else 1.0)
return list_index[s < m], data[s < m]
def __call__(self, image):
"""image (np.ndarray): BGR image"""
# Crop center image to increase speed of text detection
image_resized = crop_image(image, crop_size=CROP_SIZE).copy() if max(image.shape) > CROP_SIZE else image.copy()
poly_box_texts = self.text_detector(image_resized)
# draw_img = vis_ocr(
# image_resized,
# poly_box_texts,
# )
# cv2.imwrite("draw_img.jpg", draw_img)
is_blank = False
# Check image is blank
if len(poly_box_texts) <= MIN_NUM_BOX_TEXT:
is_blank = True
return image, is_blank, 0
# # Crop document
# poly_np = np.array(poly_box_texts)
# min_x = poly_box_texts[:, 0].min()
# max_x = poly_box_texts[:, 2].max()
# min_y = poly_box_texts[:, 1].min()
# max_y = poly_box_texts[:, 3].max()
# Filter small poly
poly_box_areas = [
[self._cal_width(poly_box), id]
for id, poly_box in enumerate(poly_box_texts)
]
poly_box_areas = sorted(poly_box_areas)[-NUMBER_BOX_FOR_ALIGNMENT:]
poly_box_areas = [poly_box_texts[id[1]] for id in poly_box_areas]
# Calculate angle
list_angle = [self._cal_angle(poly_box) for poly_box in poly_box_areas]
list_angle = [angle if angle >= MIN_ANGLE else 180 for angle in list_angle]
# LOGGER.info(f"List angle before reject outlier: {list_angle}")
list_angle = np.array(list_angle)
list_index, list_angle = self._reject_outliers(list_angle)
# LOGGER.info(f"List angle after reject outlier: {list_angle}")
if len(list_angle):
frequent_angle = self._get_most_frequent(list_angle)
list_angle = [angle for angle in list_angle if abs(angle - frequent_angle) <= 45]
# LOGGER.info(f"List angle after reject angle: {list_angle}")
angle = np.mean(list_angle)
else:
angle = 0
# LOGGER.info(f"Avg angle: {angle}")
# Reuse poly boxes detected by text detection
polys_org = PolygonsOnImage(
[Polygon(poly_box_areas[index]) for index in list_index],
shape=image_resized.shape,
)
seq_augment = iaa.Sequential([iaa.Rotate(angle, fit_output=True, order=3)])
# Rotate image by degree
if angle >= MIN_ANGLE and angle <= MAX_ANGLE:
image_resized, polys_aug = seq_augment(
image=image_resized, polygons=polys_org
)
else:
angle = 0
image_resized, polys_aug = image_resized, polys_org
# cv2.imwrite("image_resized.jpg", image_resized)
# Classify image 0 or 180 degree
list_poly = [poly.coords for poly in polys_aug]
image_crop_list = [
dewarp_by_polygon(image_resized, poly)[0] for poly in list_poly
]
cls_res = self.text_cls(image_crop_list)
cls_labels = [cls_[0] for cls_ in cls_res[1]]
# LOGGER.info(f"Angle lines: {cls_labels}")
counter = collections.Counter(cls_labels)
angle_align = angle
if counter["0"] <= counter["180"]:
aug = iaa.Rotate(angle + 180, fit_output=True, order=3)
angle_align = angle + 180
else:
aug = iaa.Rotate(angle, fit_output=True, order=3)
# Rotate the image by degree
image = aug.augment_image(image)
return image, is_blank, angle_align
# return image

View File

@ -1,41 +0,0 @@
import yaml
import pprint
import os
import json
def load_from_yaml(fname):
with open(fname, encoding='utf-8') as f:
base_config = yaml.safe_load(f)
return base_config
def load_from_json(fname):
with open(fname, "r", encoding='utf-8') as f:
base_config = json.load(f)
return base_config
class Cfg(dict):
def __init__(self, config_dict):
super(Cfg, self).__init__(**config_dict)
self.__dict__ = self
@staticmethod
def load_config_from_file(fname, download_base=False):
if not os.path.exists(fname):
raise FileNotFoundError("Not found config at {}".format(fname))
if fname.endswith(".yaml") or fname.endswith(".yml"):
return Cfg(load_from_yaml(fname))
elif fname.endswith(".json"):
return Cfg(load_from_json(fname))
else:
raise Exception(f"{fname} not supported")
def save(self, fname):
with open(fname, 'w', encoding='utf-8') as outfile:
yaml.dump(dict(self), outfile, default_flow_style=False, allow_unicode=True)
# @property
def pretty_text(self):
return pprint.PrettyPrinter().pprint(self)

View File

@ -1,75 +0,0 @@
import os
import shutil
import hashlib
import warnings
def sha256sum(filename):
h = hashlib.sha256()
b = bytearray(128*1024)
mv = memoryview(b)
with open(filename, 'rb', buffering=0) as f:
for n in iter(lambda : f.readinto(mv), 0):
h.update(mv[:n])
return h.hexdigest()
online_model_factory = {
'yolox-s-general-text-pretrain-20221226': {
'url': 'https://github.com/moewiee/satrn-model-factory/raw/main/62j266xm8r.pth',
'hash': '89bff792685af454d0cfea5d6d673be6914d614e4c2044e786da6eddf36f8b50'},
'yolox-s-checkbox-20220726': {
'url': 'https://github.com/moewiee/satrn-model-factory/raw/main/1647d7eys7.pth',
'hash': '7c1e188b7375dcf0b7b9d317675ebd92a86fdc29363558002249867249ee10f8'},
'yolox-s-idcard-5c-20221027': {
'url': 'https://github.com/moewiee/satrn-model-factory/raw/main/jr0egad3ix.pth',
'hash': '73a7772594c1f6d3f6d6a98b6d6e4097af5026864e3bd50531ad9e635ae795a7'},
'yolox-s-handwritten-text-line-20230228': {
'url': 'https://github.com/moewiee/satrn-model-factory/raw/main/rb07rtwmgi.pth',
'hash': 'a31d1bf8fc880479d2e11463dad0b4081952a13e553a02919109b634a1190ef1'}
}
__hub_available_versions__ = online_model_factory.keys()
def _get_from_hub(file_path, version, version_url):
os.system(f'wget -O {file_path} {version_url}')
assert os.path.exists(file_path), \
'wget failed while trying to retrieve from hub.'
downloaded_hash = sha256sum(file_path)
if downloaded_hash != online_model_factory[version]['hash']:
os.remove(file_path)
raise ValueError('sha256 hash doesnt match for version retrieved from hub.')
def _get(version):
use_online = version in __hub_available_versions__
if not use_online and not os.path.exists(version):
raise ValueError(f'Model version {version} not found online and not found local.')
hub_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), 'hub')
if not os.path.exists(hub_path):
os.makedirs(hub_path)
if use_online:
version_url = online_model_factory[version]['url']
file_path = os.path.join(hub_path, os.path.basename(version_url))
else:
file_path = os.path.join(hub_path, os.path.basename(version))
if not os.path.exists(file_path):
if use_online:
_get_from_hub(file_path, version, version_url)
else:
shutil.copy2(version, file_path)
else:
if use_online:
downloaded_hash = sha256sum(file_path)
if downloaded_hash != online_model_factory[version]['hash']:
os.remove(file_path)
warnings.warn('existing hub version sha256 hash doesnt match, now re-download from hub.')
_get_from_hub(file_path, version, version_url)
else:
if sha256sum(file_path) != sha256sum(version):
os.remove(file_path)
warnings.warn('existing local version sha256 hash doesnt match, now replace with new local version.')
shutil.copy2(version, file_path)
return

View File

@ -1,73 +0,0 @@
from paddleocr.tools.infer.predict_det import TextDetector
from paddleocr.tools.infer.predict_cls import TextClassifier
from paddleocr.paddleocr import parse_args
from sdsv_dewarp.config import Cfg
class PaddleTextDetector(object):
def __init__(
self,
# config_path: str,
config: dict,
use_gpu=False
):
# config = Cfg.load_config_from_file(config_path)
self.args = parse_args(mMain=False)
self.args.__dict__.update(
det_model_dir=config['model_dir'],
gpu_mem=config['gpu_mem'],
use_gpu=use_gpu,
use_zero_copy_run=True,
max_batch_size=1,
det_limit_side_len=config['det_limit_side_len'], #960
det_limit_type=config['det_limit_type'], #'max'
det_db_unclip_ratio=config['det_db_unclip_ratio'],
det_db_thresh=config['det_db_thresh'],
det_db_box_thresh=config['det_db_box_thresh'],
det_db_score_mode=config['det_db_score_mode'],
)
self.text_detector = TextDetector(self.args)
def __call__(self, image):
"""
Args:
image (np.ndarray): BGR images
Returns:
np.ndarray: numpy array of poly boxes - shape 4x2
"""
dt_boxes, time_infer = self.text_detector(image)
return dt_boxes
class PaddleTextClassifier(object):
def __init__(
self,
# config_path: str,
config: str,
use_gpu=False
):
# config = Cfg.load_config_from_file(config_path)
self.args = parse_args(mMain=False)
self.args.__dict__.update(
cls_model_dir=config['model_dir'],
gpu_mem=config['gpu_mem'],
use_gpu=use_gpu,
use_zero_copy_run=True,
cls_batch_num=config['max_batch_size'],
)
self.text_classifier = TextClassifier(self.args)
def __call__(self, images):
"""
Args:
images (np.ndarray): list of BGR images
Returns:
img_list, cls_res, elapse : cls_res format = (label, conf)
"""
out= self.text_classifier(images)
return out

View File

@ -1,212 +0,0 @@
import math
import cv2
import numpy as np
from PIL import Image, ImageDraw, ImageFont
import random
def distance(p1, p2):
"""Calculate Euclid distance"""
x1, y1 = p1
x2, y2 = p2
dist = math.sqrt((x2 - x1) ** 2 + (y2 - y1) ** 2)
return dist
def crop_image(image, crop_size=1280):
"""Crop center image"""
h, w = image.shape[:2]
x_center, y_center = w // 2, h // 2
half_size = crop_size // 2
xmin, ymin = x_center - half_size, y_center - half_size
xmax, ymax = x_center + half_size, y_center + half_size
xmin = max(xmin, 0)
ymin = max(ymin, 0)
xmax = min(xmax, w)
ymax = min(ymax, h)
return image[ymin:ymax, xmin:xmax]
def _closest_point(corners, A):
"""Find closest A in corrers point"""
distances = [distance(A, p) for p in corners]
return corners[np.argmin(distances)]
def _re_order_corners(image_size, corners) -> list:
"""Order by corners by clockwise angle"""
h, w = image_size
tl = _closest_point(corners, (0, 0))
tr = _closest_point(corners, (w, 0))
br = _closest_point(corners, (w, h))
bl = _closest_point(corners, (0, h))
return [tl, tr, br, bl]
def _validate_corner(corners, ratio_thres=0.5, epsilon=1e-3) -> bool:
"""Check corners is valid
Invalid: 3 points, duplicate points, ....
"""
c_tl, c_tr, c_br, c_bl = corners
e_top = distance(c_tl, c_tr)
e_right = distance(c_tr, c_br)
e_bottom = distance(c_br, c_bl)
e_left = distance(c_bl, c_tl)
min_tb = min(e_top, e_bottom)
max_tb = max(e_top, e_bottom)
min_lr = min(e_left, e_right)
max_lr = max(e_left, e_right)
# Nếu các điểm trùng nhau thì độ dài các cạnh sẽ bằng 0
if min(max_tb, max_lr) < epsilon:
return False
ratio = min(min_tb / max_tb, min_lr / max_lr)
if ratio < ratio_thres:
return False
return True
def dewarp_by_polygon(
image, corners, need_validate=False, need_reorder=True, trace_trans=None
):
"""Crop and dewarp from 4 corners of images
Args:
image (np.array)
corners (list): Ex : [(3347, 512), (3379, 2427), (638, 2524), (647, 495)]
need_validate (bool, optional): validate 4 points. Defaults to False.
need_reorder (bool, optional): validate 4 points. Defaults to True.
Returns:
dewarped: image after dewarp
corners: location of 4 corners after reorder
"""
h, w = image.shape[:2]
if need_reorder:
corners = _re_order_corners((h, w), corners)
dewarped = image
if need_validate:
validate = _validate_corner(corners)
else:
validate = True
if validate:
# perform dewarp
target_w = int(
max(distance(corners[0], corners[1]), distance(corners[2], corners[3]))
)
target_h = int(
max(distance(corners[0], corners[3]), distance(corners[1], corners[2]))
)
target_corners = [
[0, 0],
[target_w, 0],
[target_w, target_h],
[0, target_h],
]
pts1 = np.float32(corners)
pts2 = np.float32(target_corners)
transform_matrix = cv2.getPerspectiveTransform(pts1, pts2)
dewarped = cv2.warpPerspective(image, transform_matrix, (target_w, target_h))
if trace_trans is not None:
trace_trans["dewarp_method"]["polygon"][
"transform_matrix"
] = transform_matrix
return (dewarped, corners, trace_trans)
def vis_ocr(image, boxes, txts=[], scores=None, drop_score=0.5):
"""
Args:
image (np.ndarray / PIL): BGR image or PIL image
boxes (list / np.ndarray): list of polygon boxes
txts (list): list of text labels
scores (list, optional): probality. Defaults to None.
drop_score (float, optional): . Defaults to 0.5.
font_path (str, optional): Path of font. Defaults to "test/fonts/latin.ttf".
Returns:
np.ndarray: BGR image
"""
if len(txts) == 0:
txts = [""] * len(boxes)
if isinstance(image, np.ndarray):
image = Image.fromarray(cv2.cvtColor(image, cv2.COLOR_BGR2RGB))
if isinstance(boxes, list):
boxes = np.array(boxes)
h, w = image.height, image.width
img_left = image.copy()
img_right = Image.new("RGB", (w, h), (255, 255, 255))
draw_left = ImageDraw.Draw(img_left)
draw_right = ImageDraw.Draw(img_right)
for idx, (box, txt) in enumerate(zip(boxes, txts)):
if scores is not None and scores[idx] < drop_score:
continue
color = (random.randint(0, 255), random.randint(0, 255), random.randint(0, 255))
draw_left.polygon(
[
box[0][0],
box[0][1],
box[1][0],
box[1][1],
box[2][0],
box[2][1],
box[3][0],
box[3][1],
],
fill=color,
)
draw_right.polygon(
[
box[0][0],
box[0][1],
box[1][0],
box[1][1],
box[2][0],
box[2][1],
box[3][0],
box[3][1],
],
outline=color,
)
box_height = math.sqrt(
(box[0][0] - box[3][0]) ** 2 + (box[0][1] - box[3][1]) ** 2
)
box_width = math.sqrt(
(box[0][0] - box[1][0]) ** 2 + (box[0][1] - box[1][1]) ** 2
)
if box_height > 2 * box_width:
font_size = max(int(box_width * 0.9), 10)
font = ImageFont.load_default()
cur_y = box[0][1]
for c in txt:
char_size = font.getsize(c)
draw_right.text((box[0][0] + 3, cur_y), c, fill=(0, 0, 0), font=font)
cur_y += char_size[1]
else:
font_size = max(int(box_height * 0.8), 10)
font = ImageFont.load_default()
draw_right.text([box[0][0], box[0][1]], txt, fill=(0, 0, 0), font=font)
img_left = Image.blend(image, img_left, 0.5)
img_show = Image.new("RGB", (w * 2, h), (255, 255, 255))
img_show.paste(img_left, (0, 0, w, h))
img_show.paste(img_right, (w, 0, w * 2, h))
img_show = cv2.cvtColor(np.array(img_show), cv2.COLOR_RGB2BGR)
return img_show

View File

@ -1,187 +0,0 @@
import os
import os.path as osp
import shutil
import sys
import warnings
from setuptools import find_packages, setup
def readme():
with open('README.md', encoding='utf-8') as f:
content = f.read()
return content
version_file = 'sdsv_dewarp/version.py'
is_windows = sys.platform == 'win32'
def add_mim_extention():
"""Add extra files that are required to support MIM into the package.
These files will be added by creating a symlink to the originals if the
package is installed in `editable` mode (e.g. pip install -e .), or by
copying from the originals otherwise.
"""
# parse installment mode
if 'develop' in sys.argv:
# installed by `pip install -e .`
mode = 'symlink'
elif 'sdist' in sys.argv or 'bdist_wheel' in sys.argv:
# installed by `pip install .`
# or create source distribution by `python setup.py sdist`
mode = 'copy'
else:
return
filenames = ['tools', 'configs', 'model-index.yml']
repo_path = osp.dirname(__file__)
mim_path = osp.join(repo_path, 'mmocr', '.mim')
os.makedirs(mim_path, exist_ok=True)
for filename in filenames:
if osp.exists(filename):
src_path = osp.join(repo_path, filename)
tar_path = osp.join(mim_path, filename)
if osp.isfile(tar_path) or osp.islink(tar_path):
os.remove(tar_path)
elif osp.isdir(tar_path):
shutil.rmtree(tar_path)
if mode == 'symlink':
src_relpath = osp.relpath(src_path, osp.dirname(tar_path))
try:
os.symlink(src_relpath, tar_path)
except OSError:
# Creating a symbolic link on windows may raise an
# `OSError: [WinError 1314]` due to privilege. If
# the error happens, the src file will be copied
mode = 'copy'
warnings.warn(
f'Failed to create a symbolic link for {src_relpath}, '
f'and it will be copied to {tar_path}')
else:
continue
if mode == 'copy':
if osp.isfile(src_path):
shutil.copyfile(src_path, tar_path)
elif osp.isdir(src_path):
shutil.copytree(src_path, tar_path)
else:
warnings.warn(f'Cannot copy file {src_path}.')
else:
raise ValueError(f'Invalid mode {mode}')
def get_version():
with open(version_file, 'r') as f:
exec(compile(f.read(), version_file, 'exec'))
import sys
# return short version for sdist
if 'sdist' in sys.argv or 'bdist_wheel' in sys.argv:
return locals()['short_version']
else:
return locals()['__version__']
def parse_requirements(fname='requirements.txt', with_version=True):
"""Parse the package dependencies listed in a requirements file but strip
specific version information.
Args:
fname (str): Path to requirements file.
with_version (bool, default=False): If True, include version specs.
Returns:
info (list[str]): List of requirements items.
CommandLine:
python -c "import setup; print(setup.parse_requirements())"
"""
import re
import sys
from os.path import exists
require_fpath = fname
def parse_line(line):
"""Parse information from a line in a requirements text file."""
if line.startswith('-r '):
# Allow specifying requirements in other files
target = line.split(' ')[1]
for info in parse_require_file(target):
yield info
else:
info = {'line': line}
if line.startswith('-e '):
info['package'] = line.split('#egg=')[1]
else:
# Remove versioning from the package
pat = '(' + '|'.join(['>=', '==', '>']) + ')'
parts = re.split(pat, line, maxsplit=1)
parts = [p.strip() for p in parts]
info['package'] = parts[0]
if len(parts) > 1:
op, rest = parts[1:]
if ';' in rest:
# Handle platform specific dependencies
# http://setuptools.readthedocs.io/en/latest/setuptools.html#declaring-platform-specific-dependencies
version, platform_deps = map(str.strip,
rest.split(';'))
info['platform_deps'] = platform_deps
else:
version = rest # NOQA
info['version'] = (op, version)
yield info
def parse_require_file(fpath):
with open(fpath, 'r') as f:
for line in f.readlines():
line = line.strip()
if line and not line.startswith('#'):
for info in parse_line(line):
yield info
def gen_packages_items():
if exists(require_fpath):
for info in parse_require_file(require_fpath):
parts = [info['package']]
if with_version and 'version' in info:
parts.extend(info['version'])
if not sys.version.startswith('3.4'):
# apparently package_deps are broken in 3.4
platform_deps = info.get('platform_deps')
if platform_deps is not None:
parts.append(';' + platform_deps)
item = ''.join(parts)
yield item
packages = list(gen_packages_items())
return packages
if __name__ == '__main__':
setup(
name='sdsv_dewarp',
version=get_version(),
description='Dewarp document',
long_description=readme(),
long_description_content_type='text/markdown',
packages=find_packages(exclude=('configs', 'tools', 'demo')),
include_package_data=True,
url='',
classifiers=[
'Development Status :: 4 - Beta',
'License :: OSI Approved :: Apache Software License',
'Operating System :: OS Independent',
'Programming Language :: Python :: 3',
'Programming Language :: Python :: 3.6',
'Programming Language :: Python :: 3.7',
'Programming Language :: Python :: 3.8',
'Programming Language :: Python :: 3.9',
],
license='Apache License 2.0',
install_requires=parse_requirements('requirements.txt'),
zip_safe=False)

View File

@ -1,47 +0,0 @@
from sdsv_dewarp.api import AlignImage
import cv2
import glob
import os
import tqdm
import time
import argparse
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--input")
parser.add_argument("--out")
parser.add_argument("--device", type=str, default="cuda:1")
args = parser.parse_args()
model = AlignImage(device=args.device)
img_dir = args.input
out_dir = args.out
if not os.path.exists(out_dir):
os.makedirs(out_dir)
img_paths = glob.glob(img_dir + "/*")
times = []
for img_path in tqdm.tqdm(img_paths):
t1 = time.time()
img = cv2.imread(img_path)
if img is None:
print(img_path)
continue
aligned_img, is_blank, angle_align = model(img)
times.append(time.time() - t1)
if not is_blank:
cv2.imwrite(os.path.join(out_dir, os.path.basename(img_path)), aligned_img)
else:
cv2.imwrite(os.path.join(out_dir, os.path.basename(img_path)), img)
times = times[1:]
print("Avg time: ", sum(times) / len(times))

View File

@ -1,6 +0,0 @@
# Builds
*.egg-info
__pycache__
# Checkpoint
hub

View File

@ -1,674 +0,0 @@
GNU GENERAL PUBLIC LICENSE
Version 3, 29 June 2007
Copyright (C) 2007 Free Software Foundation, Inc. <http://fsf.org/>
Everyone is permitted to copy and distribute verbatim copies
of this license document, but changing it is not allowed.
Preamble
The GNU General Public License is a free, copyleft license for
software and other kinds of works.
The licenses for most software and other practical works are designed
to take away your freedom to share and change the works. By contrast,
the GNU General Public License is intended to guarantee your freedom to
share and change all versions of a program--to make sure it remains free
software for all its users. We, the Free Software Foundation, use the
GNU General Public License for most of our software; it applies also to
any other work released this way by its authors. You can apply it to
your programs, too.
When we speak of free software, we are referring to freedom, not
price. Our General Public Licenses are designed to make sure that you
have the freedom to distribute copies of free software (and charge for
them if you wish), that you receive source code or can get it if you
want it, that you can change the software or use pieces of it in new
free programs, and that you know you can do these things.
To protect your rights, we need to prevent others from denying you
these rights or asking you to surrender the rights. Therefore, you have
certain responsibilities if you distribute copies of the software, or if
you modify it: responsibilities to respect the freedom of others.
For example, if you distribute copies of such a program, whether
gratis or for a fee, you must pass on to the recipients the same
freedoms that you received. You must make sure that they, too, receive
or can get the source code. And you must show them these terms so they
know their rights.
Developers that use the GNU GPL protect your rights with two steps:
(1) assert copyright on the software, and (2) offer you this License
giving you legal permission to copy, distribute and/or modify it.
For the developers' and authors' protection, the GPL clearly explains
that there is no warranty for this free software. For both users' and
authors' sake, the GPL requires that modified versions be marked as
changed, so that their problems will not be attributed erroneously to
authors of previous versions.
Some devices are designed to deny users access to install or run
modified versions of the software inside them, although the manufacturer
can do so. This is fundamentally incompatible with the aim of
protecting users' freedom to change the software. The systematic
pattern of such abuse occurs in the area of products for individuals to
use, which is precisely where it is most unacceptable. Therefore, we
have designed this version of the GPL to prohibit the practice for those
products. If such problems arise substantially in other domains, we
stand ready to extend this provision to those domains in future versions
of the GPL, as needed to protect the freedom of users.
Finally, every program is threatened constantly by software patents.
States should not allow patents to restrict development and use of
software on general-purpose computers, but in those that do, we wish to
avoid the special danger that patents applied to a free program could
make it effectively proprietary. To prevent this, the GPL assures that
patents cannot be used to render the program non-free.
The precise terms and conditions for copying, distribution and
modification follow.
TERMS AND CONDITIONS
0. Definitions.
"This License" refers to version 3 of the GNU General Public License.
"Copyright" also means copyright-like laws that apply to other kinds of
works, such as semiconductor masks.
"The Program" refers to any copyrightable work licensed under this
License. Each licensee is addressed as "you". "Licensees" and
"recipients" may be individuals or organizations.
To "modify" a work means to copy from or adapt all or part of the work
in a fashion requiring copyright permission, other than the making of an
exact copy. The resulting work is called a "modified version" of the
earlier work or a work "based on" the earlier work.
A "covered work" means either the unmodified Program or a work based
on the Program.
To "propagate" a work means to do anything with it that, without
permission, would make you directly or secondarily liable for
infringement under applicable copyright law, except executing it on a
computer or modifying a private copy. Propagation includes copying,
distribution (with or without modification), making available to the
public, and in some countries other activities as well.
To "convey" a work means any kind of propagation that enables other
parties to make or receive copies. Mere interaction with a user through
a computer network, with no transfer of a copy, is not conveying.
An interactive user interface displays "Appropriate Legal Notices"
to the extent that it includes a convenient and prominently visible
feature that (1) displays an appropriate copyright notice, and (2)
tells the user that there is no warranty for the work (except to the
extent that warranties are provided), that licensees may convey the
work under this License, and how to view a copy of this License. If
the interface presents a list of user commands or options, such as a
menu, a prominent item in the list meets this criterion.
1. Source Code.
The "source code" for a work means the preferred form of the work
for making modifications to it. "Object code" means any non-source
form of a work.
A "Standard Interface" means an interface that either is an official
standard defined by a recognized standards body, or, in the case of
interfaces specified for a particular programming language, one that
is widely used among developers working in that language.
The "System Libraries" of an executable work include anything, other
than the work as a whole, that (a) is included in the normal form of
packaging a Major Component, but which is not part of that Major
Component, and (b) serves only to enable use of the work with that
Major Component, or to implement a Standard Interface for which an
implementation is available to the public in source code form. A
"Major Component", in this context, means a major essential component
(kernel, window system, and so on) of the specific operating system
(if any) on which the executable work runs, or a compiler used to
produce the work, or an object code interpreter used to run it.
The "Corresponding Source" for a work in object code form means all
the source code needed to generate, install, and (for an executable
work) run the object code and to modify the work, including scripts to
control those activities. However, it does not include the work's
System Libraries, or general-purpose tools or generally available free
programs which are used unmodified in performing those activities but
which are not part of the work. For example, Corresponding Source
includes interface definition files associated with source files for
the work, and the source code for shared libraries and dynamically
linked subprograms that the work is specifically designed to require,
such as by intimate data communication or control flow between those
subprograms and other parts of the work.
The Corresponding Source need not include anything that users
can regenerate automatically from other parts of the Corresponding
Source.
The Corresponding Source for a work in source code form is that
same work.
2. Basic Permissions.
All rights granted under this License are granted for the term of
copyright on the Program, and are irrevocable provided the stated
conditions are met. This License explicitly affirms your unlimited
permission to run the unmodified Program. The output from running a
covered work is covered by this License only if the output, given its
content, constitutes a covered work. This License acknowledges your
rights of fair use or other equivalent, as provided by copyright law.
You may make, run and propagate covered works that you do not
convey, without conditions so long as your license otherwise remains
in force. You may convey covered works to others for the sole purpose
of having them make modifications exclusively for you, or provide you
with facilities for running those works, provided that you comply with
the terms of this License in conveying all material for which you do
not control copyright. Those thus making or running the covered works
for you must do so exclusively on your behalf, under your direction
and control, on terms that prohibit them from making any copies of
your copyrighted material outside their relationship with you.
Conveying under any other circumstances is permitted solely under
the conditions stated below. Sublicensing is not allowed; section 10
makes it unnecessary.
3. Protecting Users' Legal Rights From Anti-Circumvention Law.
No covered work shall be deemed part of an effective technological
measure under any applicable law fulfilling obligations under article
11 of the WIPO copyright treaty adopted on 20 December 1996, or
similar laws prohibiting or restricting circumvention of such
measures.
When you convey a covered work, you waive any legal power to forbid
circumvention of technological measures to the extent such circumvention
is effected by exercising rights under this License with respect to
the covered work, and you disclaim any intention to limit operation or
modification of the work as a means of enforcing, against the work's
users, your or third parties' legal rights to forbid circumvention of
technological measures.
4. Conveying Verbatim Copies.
You may convey verbatim copies of the Program's source code as you
receive it, in any medium, provided that you conspicuously and
appropriately publish on each copy an appropriate copyright notice;
keep intact all notices stating that this License and any
non-permissive terms added in accord with section 7 apply to the code;
keep intact all notices of the absence of any warranty; and give all
recipients a copy of this License along with the Program.
You may charge any price or no price for each copy that you convey,
and you may offer support or warranty protection for a fee.
5. Conveying Modified Source Versions.
You may convey a work based on the Program, or the modifications to
produce it from the Program, in the form of source code under the
terms of section 4, provided that you also meet all of these conditions:
a) The work must carry prominent notices stating that you modified
it, and giving a relevant date.
b) The work must carry prominent notices stating that it is
released under this License and any conditions added under section
7. This requirement modifies the requirement in section 4 to
"keep intact all notices".
c) You must license the entire work, as a whole, under this
License to anyone who comes into possession of a copy. This
License will therefore apply, along with any applicable section 7
additional terms, to the whole of the work, and all its parts,
regardless of how they are packaged. This License gives no
permission to license the work in any other way, but it does not
invalidate such permission if you have separately received it.
d) If the work has interactive user interfaces, each must display
Appropriate Legal Notices; however, if the Program has interactive
interfaces that do not display Appropriate Legal Notices, your
work need not make them do so.
A compilation of a covered work with other separate and independent
works, which are not by their nature extensions of the covered work,
and which are not combined with it such as to form a larger program,
in or on a volume of a storage or distribution medium, is called an
"aggregate" if the compilation and its resulting copyright are not
used to limit the access or legal rights of the compilation's users
beyond what the individual works permit. Inclusion of a covered work
in an aggregate does not cause this License to apply to the other
parts of the aggregate.
6. Conveying Non-Source Forms.
You may convey a covered work in object code form under the terms
of sections 4 and 5, provided that you also convey the
machine-readable Corresponding Source under the terms of this License,
in one of these ways:
a) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by the
Corresponding Source fixed on a durable physical medium
customarily used for software interchange.
b) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by a
written offer, valid for at least three years and valid for as
long as you offer spare parts or customer support for that product
model, to give anyone who possesses the object code either (1) a
copy of the Corresponding Source for all the software in the
product that is covered by this License, on a durable physical
medium customarily used for software interchange, for a price no
more than your reasonable cost of physically performing this
conveying of source, or (2) access to copy the
Corresponding Source from a network server at no charge.
c) Convey individual copies of the object code with a copy of the
written offer to provide the Corresponding Source. This
alternative is allowed only occasionally and noncommercially, and
only if you received the object code with such an offer, in accord
with subsection 6b.
d) Convey the object code by offering access from a designated
place (gratis or for a charge), and offer equivalent access to the
Corresponding Source in the same way through the same place at no
further charge. You need not require recipients to copy the
Corresponding Source along with the object code. If the place to
copy the object code is a network server, the Corresponding Source
may be on a different server (operated by you or a third party)
that supports equivalent copying facilities, provided you maintain
clear directions next to the object code saying where to find the
Corresponding Source. Regardless of what server hosts the
Corresponding Source, you remain obligated to ensure that it is
available for as long as needed to satisfy these requirements.
e) Convey the object code using peer-to-peer transmission, provided
you inform other peers where the object code and Corresponding
Source of the work are being offered to the general public at no
charge under subsection 6d.
A separable portion of the object code, whose source code is excluded
from the Corresponding Source as a System Library, need not be
included in conveying the object code work.
A "User Product" is either (1) a "consumer product", which means any
tangible personal property which is normally used for personal, family,
or household purposes, or (2) anything designed or sold for incorporation
into a dwelling. In determining whether a product is a consumer product,
doubtful cases shall be resolved in favor of coverage. For a particular
product received by a particular user, "normally used" refers to a
typical or common use of that class of product, regardless of the status
of the particular user or of the way in which the particular user
actually uses, or expects or is expected to use, the product. A product
is a consumer product regardless of whether the product has substantial
commercial, industrial or non-consumer uses, unless such uses represent
the only significant mode of use of the product.
"Installation Information" for a User Product means any methods,
procedures, authorization keys, or other information required to install
and execute modified versions of a covered work in that User Product from
a modified version of its Corresponding Source. The information must
suffice to ensure that the continued functioning of the modified object
code is in no case prevented or interfered with solely because
modification has been made.
If you convey an object code work under this section in, or with, or
specifically for use in, a User Product, and the conveying occurs as
part of a transaction in which the right of possession and use of the
User Product is transferred to the recipient in perpetuity or for a
fixed term (regardless of how the transaction is characterized), the
Corresponding Source conveyed under this section must be accompanied
by the Installation Information. But this requirement does not apply
if neither you nor any third party retains the ability to install
modified object code on the User Product (for example, the work has
been installed in ROM).
The requirement to provide Installation Information does not include a
requirement to continue to provide support service, warranty, or updates
for a work that has been modified or installed by the recipient, or for
the User Product in which it has been modified or installed. Access to a
network may be denied when the modification itself materially and
adversely affects the operation of the network or violates the rules and
protocols for communication across the network.
Corresponding Source conveyed, and Installation Information provided,
in accord with this section must be in a format that is publicly
documented (and with an implementation available to the public in
source code form), and must require no special password or key for
unpacking, reading or copying.
7. Additional Terms.
"Additional permissions" are terms that supplement the terms of this
License by making exceptions from one or more of its conditions.
Additional permissions that are applicable to the entire Program shall
be treated as though they were included in this License, to the extent
that they are valid under applicable law. If additional permissions
apply only to part of the Program, that part may be used separately
under those permissions, but the entire Program remains governed by
this License without regard to the additional permissions.
When you convey a copy of a covered work, you may at your option
remove any additional permissions from that copy, or from any part of
it. (Additional permissions may be written to require their own
removal in certain cases when you modify the work.) You may place
additional permissions on material, added by you to a covered work,
for which you have or can give appropriate copyright permission.
Notwithstanding any other provision of this License, for material you
add to a covered work, you may (if authorized by the copyright holders of
that material) supplement the terms of this License with terms:
a) Disclaiming warranty or limiting liability differently from the
terms of sections 15 and 16 of this License; or
b) Requiring preservation of specified reasonable legal notices or
author attributions in that material or in the Appropriate Legal
Notices displayed by works containing it; or
c) Prohibiting misrepresentation of the origin of that material, or
requiring that modified versions of such material be marked in
reasonable ways as different from the original version; or
d) Limiting the use for publicity purposes of names of licensors or
authors of the material; or
e) Declining to grant rights under trademark law for use of some
trade names, trademarks, or service marks; or
f) Requiring indemnification of licensors and authors of that
material by anyone who conveys the material (or modified versions of
it) with contractual assumptions of liability to the recipient, for
any liability that these contractual assumptions directly impose on
those licensors and authors.
All other non-permissive additional terms are considered "further
restrictions" within the meaning of section 10. If the Program as you
received it, or any part of it, contains a notice stating that it is
governed by this License along with a term that is a further
restriction, you may remove that term. If a license document contains
a further restriction but permits relicensing or conveying under this
License, you may add to a covered work material governed by the terms
of that license document, provided that the further restriction does
not survive such relicensing or conveying.
If you add terms to a covered work in accord with this section, you
must place, in the relevant source files, a statement of the
additional terms that apply to those files, or a notice indicating
where to find the applicable terms.
Additional terms, permissive or non-permissive, may be stated in the
form of a separately written license, or stated as exceptions;
the above requirements apply either way.
8. Termination.
You may not propagate or modify a covered work except as expressly
provided under this License. Any attempt otherwise to propagate or
modify it is void, and will automatically terminate your rights under
this License (including any patent licenses granted under the third
paragraph of section 11).
However, if you cease all violation of this License, then your
license from a particular copyright holder is reinstated (a)
provisionally, unless and until the copyright holder explicitly and
finally terminates your license, and (b) permanently, if the copyright
holder fails to notify you of the violation by some reasonable means
prior to 60 days after the cessation.
Moreover, your license from a particular copyright holder is
reinstated permanently if the copyright holder notifies you of the
violation by some reasonable means, this is the first time you have
received notice of violation of this License (for any work) from that
copyright holder, and you cure the violation prior to 30 days after
your receipt of the notice.
Termination of your rights under this section does not terminate the
licenses of parties who have received copies or rights from you under
this License. If your rights have been terminated and not permanently
reinstated, you do not qualify to receive new licenses for the same
material under section 10.
9. Acceptance Not Required for Having Copies.
You are not required to accept this License in order to receive or
run a copy of the Program. Ancillary propagation of a covered work
occurring solely as a consequence of using peer-to-peer transmission
to receive a copy likewise does not require acceptance. However,
nothing other than this License grants you permission to propagate or
modify any covered work. These actions infringe copyright if you do
not accept this License. Therefore, by modifying or propagating a
covered work, you indicate your acceptance of this License to do so.
10. Automatic Licensing of Downstream Recipients.
Each time you convey a covered work, the recipient automatically
receives a license from the original licensors, to run, modify and
propagate that work, subject to this License. You are not responsible
for enforcing compliance by third parties with this License.
An "entity transaction" is a transaction transferring control of an
organization, or substantially all assets of one, or subdividing an
organization, or merging organizations. If propagation of a covered
work results from an entity transaction, each party to that
transaction who receives a copy of the work also receives whatever
licenses to the work the party's predecessor in interest had or could
give under the previous paragraph, plus a right to possession of the
Corresponding Source of the work from the predecessor in interest, if
the predecessor has it or can get it with reasonable efforts.
You may not impose any further restrictions on the exercise of the
rights granted or affirmed under this License. For example, you may
not impose a license fee, royalty, or other charge for exercise of
rights granted under this License, and you may not initiate litigation
(including a cross-claim or counterclaim in a lawsuit) alleging that
any patent claim is infringed by making, using, selling, offering for
sale, or importing the Program or any portion of it.
11. Patents.
A "contributor" is a copyright holder who authorizes use under this
License of the Program or a work on which the Program is based. The
work thus licensed is called the contributor's "contributor version".
A contributor's "essential patent claims" are all patent claims
owned or controlled by the contributor, whether already acquired or
hereafter acquired, that would be infringed by some manner, permitted
by this License, of making, using, or selling its contributor version,
but do not include claims that would be infringed only as a
consequence of further modification of the contributor version. For
purposes of this definition, "control" includes the right to grant
patent sublicenses in a manner consistent with the requirements of
this License.
Each contributor grants you a non-exclusive, worldwide, royalty-free
patent license under the contributor's essential patent claims, to
make, use, sell, offer for sale, import and otherwise run, modify and
propagate the contents of its contributor version.
In the following three paragraphs, a "patent license" is any express
agreement or commitment, however denominated, not to enforce a patent
(such as an express permission to practice a patent or covenant not to
sue for patent infringement). To "grant" such a patent license to a
party means to make such an agreement or commitment not to enforce a
patent against the party.
If you convey a covered work, knowingly relying on a patent license,
and the Corresponding Source of the work is not available for anyone
to copy, free of charge and under the terms of this License, through a
publicly available network server or other readily accessible means,
then you must either (1) cause the Corresponding Source to be so
available, or (2) arrange to deprive yourself of the benefit of the
patent license for this particular work, or (3) arrange, in a manner
consistent with the requirements of this License, to extend the patent
license to downstream recipients. "Knowingly relying" means you have
actual knowledge that, but for the patent license, your conveying the
covered work in a country, or your recipient's use of the covered work
in a country, would infringe one or more identifiable patents in that
country that you have reason to believe are valid.
If, pursuant to or in connection with a single transaction or
arrangement, you convey, or propagate by procuring conveyance of, a
covered work, and grant a patent license to some of the parties
receiving the covered work authorizing them to use, propagate, modify
or convey a specific copy of the covered work, then the patent license
you grant is automatically extended to all recipients of the covered
work and works based on it.
A patent license is "discriminatory" if it does not include within
the scope of its coverage, prohibits the exercise of, or is
conditioned on the non-exercise of one or more of the rights that are
specifically granted under this License. You may not convey a covered
work if you are a party to an arrangement with a third party that is
in the business of distributing software, under which you make payment
to the third party based on the extent of your activity of conveying
the work, and under which the third party grants, to any of the
parties who would receive the covered work from you, a discriminatory
patent license (a) in connection with copies of the covered work
conveyed by you (or copies made from those copies), or (b) primarily
for and in connection with specific products or compilations that
contain the covered work, unless you entered into that arrangement,
or that patent license was granted, prior to 28 March 2007.
Nothing in this License shall be construed as excluding or limiting
any implied license or other defenses to infringement that may
otherwise be available to you under applicable patent law.
12. No Surrender of Others' Freedom.
If conditions are imposed on you (whether by court order, agreement or
otherwise) that contradict the conditions of this License, they do not
excuse you from the conditions of this License. If you cannot convey a
covered work so as to satisfy simultaneously your obligations under this
License and any other pertinent obligations, then as a consequence you may
not convey it at all. For example, if you agree to terms that obligate you
to collect a royalty for further conveying from those to whom you convey
the Program, the only way you could satisfy both those terms and this
License would be to refrain entirely from conveying the Program.
13. Use with the GNU Affero General Public License.
Notwithstanding any other provision of this License, you have
permission to link or combine any covered work with a work licensed
under version 3 of the GNU Affero General Public License into a single
combined work, and to convey the resulting work. The terms of this
License will continue to apply to the part which is the covered work,
but the special requirements of the GNU Affero General Public License,
section 13, concerning interaction through a network will apply to the
combination as such.
14. Revised Versions of this License.
The Free Software Foundation may publish revised and/or new versions of
the GNU General Public License from time to time. Such new versions will
be similar in spirit to the present version, but may differ in detail to
address new problems or concerns.
Each version is given a distinguishing version number. If the
Program specifies that a certain numbered version of the GNU General
Public License "or any later version" applies to it, you have the
option of following the terms and conditions either of that numbered
version or of any later version published by the Free Software
Foundation. If the Program does not specify a version number of the
GNU General Public License, you may choose any version ever published
by the Free Software Foundation.
If the Program specifies that a proxy can decide which future
versions of the GNU General Public License can be used, that proxy's
public statement of acceptance of a version permanently authorizes you
to choose that version for the Program.
Later license versions may give you additional or different
permissions. However, no additional obligations are imposed on any
author or copyright holder as a result of your choosing to follow a
later version.
15. Disclaimer of Warranty.
THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
16. Limitation of Liability.
IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
SUCH DAMAGES.
17. Interpretation of Sections 15 and 16.
If the disclaimer of warranty and limitation of liability provided
above cannot be given local legal effect according to their terms,
reviewing courts shall apply local law that most closely approximates
an absolute waiver of all civil liability in connection with the
Program, unless a warranty or assumption of liability accompanies a
copy of the Program in return for a fee.
END OF TERMS AND CONDITIONS
How to Apply These Terms to Your New Programs
If you develop a new program, and you want it to be of the greatest
possible use to the public, the best way to achieve this is to make it
free software which everyone can redistribute and change under these terms.
To do so, attach the following notices to the program. It is safest
to attach them to the start of each source file to most effectively
state the exclusion of warranty; and each file should have at least
the "copyright" line and a pointer to where the full notice is found.
<one line to give the program's name and a brief idea of what it does.>
Copyright (C) <year> <name of author>
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see <http://www.gnu.org/licenses/>.
Also add information on how to contact you by electronic and paper mail.
If the program does terminal interaction, make it output a short
notice like this when it starts in an interactive mode:
<program> Copyright (C) <year> <name of author>
This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
This is free software, and you are welcome to redistribute it
under certain conditions; type `show c' for details.
The hypothetical commands `show w' and `show c' should show the appropriate
parts of the General Public License. Of course, your program's commands
might be different; for a GUI interface, you would use an "about box".
You should also get your employer (if you work as a programmer) or school,
if any, to sign a "copyright disclaimer" for the program, if necessary.
For more information on this, and how to apply and follow the GNU GPL, see
<http://www.gnu.org/licenses/>.
The GNU General Public License does not permit incorporating your program
into proprietary programs. If your program is a subroutine library, you
may consider it more useful to permit linking proprietary applications with
the library. If this is what you want to do, use the GNU Lesser General
Public License instead of this License. But first, please read
<http://www.gnu.org/philosophy/why-not-lgpl.html>.

View File

@ -1,76 +0,0 @@
## Introduction
This repo serve as source code storage for the Standalone YoloX Detection packages.
Installing this package requires 2 additional packages: PyTorch and MMCV.
## Installation
```shell
conda create -n sdsvtd-env python=3.8
conda activate sdsvtd-env
conda install pytorch torchvision pytorch-cuda=11.6 -c pytorch -c nvidia
pip install mmcv-full
git clone https://github.com/moewiee/sdsvtd.git
cd sdsvtd
pip install -v -e .
```
## Basic Usage
```python
from sdsvtd import StandaloneYOLOXRunner
runner = StandaloneYOLOXRunner(version='yolox-s-general-text-pretrain-20221226', device='cpu')
```
The `version` parameter accepts version names declared in `sdsvtd.factory.online_model_factory` or a local path such as `$DIR\model.pth`. To check for available versions in the hub, run:
```python
import sdsvtd
print(sdsvtd.__hub_available_versions__)
```
Naturally, a `StandaloneYOLOXRunner` instance assumes the input to be an instance of `np.ndarray` or an instance of `str` (batch inferece is not supported), for examples:
```python
import numpy as np
from sdsvtd import StandaloneYOLOXRunner
runner = StandaloneYOLOXRunner(version='yolox-s-general-text-pretrain-20221226', device='cpu')
dummy_input = np.ndarray(500, 500, 3)
result = runner(dummy_input)
```
**Note:** Output of `StandaloneYOLOXRunner` instance will be in format `List[np.ndarray]` with each list element corresponds to one class. Each `np.ndarray` will be a 5-d vector `[x1 y1 x2 y2 confident]` with coordinates rescaled to fit the original image size.
**AutoRotation:**
From version 0.1.0, `sdsvtd` support *AutoRotation* feature which accept a `np.ndarray/str` as input and return an straight rotated image (only available rotation degrees are 90, 180 and 270) of type `np.ndarray` and its bounding boxes of type `List[np.ndarray]`. Usage:
```python
import numpy as np
from sdsvtd import StandaloneYOLOXRunner
runner = StandaloneYOLOXRunner(version='yolox-s-general-text-pretrain-20221226', device='cpu', auto_rotate=True)
rotated_image, result = runner(cv2.imread('path-to-image')) # or
rotated_image, result = runner(np.ndarray)
```
## Version Changelog
* **[0.0.1]**
Initial version with specified features.
* **[0.0.2]**
Update more versions in model hub.
* **[0.0.3]**
Update feature to specify running device while initialize runner. [Issue](https://github.com/moewiee/sdsvtd/issues/2)
* **[0.0.4]**
Fix a minor bugs when existing hub/local version != current hub/local version.
* **[0.0.5]**
Update model hub with handwritten text line detection version.
* **[0.1.0]**
Update new feature: Auto Rotation.
* **[0.1.1]**
Fix a bug in API inference class where return double result with auto_rotate=False.
* **[0.1.2]**
Fix a bug in rotate feature where rotator_version was ignored.

View File

@ -1,3 +0,0 @@
from .api import StandaloneYOLOXRunner
from .version import __version__
from .factory import __hub_available_versions__

View File

@ -1,35 +0,0 @@
from .model import SingleStageDetector, AutoRotateDetector
import numpy as np
class StandaloneYOLOXRunner:
def __init__(self,
version,
device,
auto_rotate=False,
rotator_version=None):
self.model = SingleStageDetector(version,
device)
self.auto_rotate = auto_rotate
if self.auto_rotate:
if rotator_version is None:
rotator_version = version
self.rotator = AutoRotateDetector(rotator_version,
device)
self.warmup_()
def warmup_(self):
''' Call on dummpy input to warm up instance '''
x = np.ndarray((1280, 1280, 3)).astype(np.uint8)
self.model(x)
if self.auto_rotate:
self.rotator(x)
def __call__(self, img):
if self.auto_rotate:
img = self.rotator(img)
result = self.model(img)
return result if not self.auto_rotate else (img, result)

View File

@ -1,395 +0,0 @@
import torch
import torch.nn as nn
from mmcv.cnn import ConvModule
from torch.nn.modules.batchnorm import _BatchNorm
class DarknetBottleneck(nn.Module):
"""The basic bottleneck block used in Darknet.
Each ResBlock consists of two ConvModules and the input is added to the
final output. Each ConvModule is composed of Conv, BN, and LeakyReLU.
The first convLayer has filter size of 1x1 and the second one has the
filter size of 3x3.
Args:
in_channels (int): The input channels of this Module.
out_channels (int): The output channels of this Module.
expansion (int): The kernel size of the convolution. Default: 0.5
add_identity (bool): Whether to add identity to the out.
Default: True
use_depthwise (bool): Whether to use depthwise separable convolution.
Default: False
conv_cfg (dict): Config dict for convolution layer. Default: None,
which means using conv2d.
norm_cfg (dict): Config dict for normalization layer.
Default: dict(type='BN').
act_cfg (dict): Config dict for activation layer.
Default: dict(type='Swish').
"""
def __init__(self,
in_channels,
out_channels,
expansion=0.5,
add_identity=True,
use_depthwise=False,
conv_cfg=None,
norm_cfg=dict(type='BN', momentum=0.03, eps=0.001),
act_cfg=dict(type='Swish')):
super().__init__()
hidden_channels = int(out_channels * expansion)
self.conv1 = ConvModule(
in_channels,
hidden_channels,
1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg)
self.conv2 = ConvModule(
hidden_channels,
out_channels,
3,
stride=1,
padding=1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg)
self.add_identity = \
add_identity and in_channels == out_channels
def forward(self, x):
identity = x
out = self.conv1(x)
out = self.conv2(out)
if self.add_identity:
return out + identity
else:
return out
class CSPLayer(nn.Module):
"""Cross Stage Partial Layer.
Args:
in_channels (int): The input channels of the CSP layer.
out_channels (int): The output channels of the CSP layer.
expand_ratio (float): Ratio to adjust the number of channels of the
hidden layer. Default: 0.5
num_blocks (int): Number of blocks. Default: 1
add_identity (bool): Whether to add identity in blocks.
Default: True
use_depthwise (bool): Whether to depthwise separable convolution in
blocks. Default: False
conv_cfg (dict, optional): Config dict for convolution layer.
Default: None, which means using conv2d.
norm_cfg (dict): Config dict for normalization layer.
Default: dict(type='BN')
act_cfg (dict): Config dict for activation layer.
Default: dict(type='Swish')
"""
def __init__(self,
in_channels,
out_channels,
expand_ratio=0.5,
num_blocks=1,
add_identity=True,
use_depthwise=False,
conv_cfg=None,
norm_cfg=dict(type='BN', momentum=0.03, eps=0.001),
act_cfg=dict(type='Swish')):
super().__init__()
mid_channels = int(out_channels * expand_ratio)
self.main_conv = ConvModule(
in_channels,
mid_channels,
1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg)
self.short_conv = ConvModule(
in_channels,
mid_channels,
1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg)
self.final_conv = ConvModule(
2 * mid_channels,
out_channels,
1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg)
self.blocks = nn.Sequential(*[
DarknetBottleneck(
mid_channels,
mid_channels,
1.0,
add_identity,
use_depthwise,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg) for _ in range(num_blocks)
])
def forward(self, x):
x_short = self.short_conv(x)
x_main = self.main_conv(x)
x_main = self.blocks(x_main)
x_final = torch.cat((x_main, x_short), dim=1)
return self.final_conv(x_final)
class Focus(nn.Module):
"""Focus width and height information into channel space.
Args:
in_channels (int): The input channels of this Module.
out_channels (int): The output channels of this Module.
kernel_size (int): The kernel size of the convolution. Default: 1
stride (int): The stride of the convolution. Default: 1
conv_cfg (dict): Config dict for convolution layer. Default: None,
which means using conv2d.
norm_cfg (dict): Config dict for normalization layer.
Default: dict(type='BN', momentum=0.03, eps=0.001).
act_cfg (dict): Config dict for activation layer.
Default: dict(type='Swish').
"""
def __init__(self,
in_channels,
out_channels,
kernel_size=1,
stride=1,
conv_cfg=None,
norm_cfg=dict(type='BN', momentum=0.03, eps=0.001),
act_cfg=dict(type='Swish')):
super().__init__()
self.conv = ConvModule(
in_channels * 4,
out_channels,
kernel_size,
stride,
padding=(kernel_size - 1) // 2,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg)
def forward(self, x):
# shape of x (b,c,w,h) -> y(b,4c,w/2,h/2)
patch_top_left = x[..., ::2, ::2]
patch_top_right = x[..., ::2, 1::2]
patch_bot_left = x[..., 1::2, ::2]
patch_bot_right = x[..., 1::2, 1::2]
x = torch.cat(
(
patch_top_left,
patch_bot_left,
patch_top_right,
patch_bot_right,
),
dim=1,
)
return self.conv(x)
class SPPBottleneck(nn.Module):
"""Spatial pyramid pooling layer used in YOLOv3-SPP.
Args:
in_channels (int): The input channels of this Module.
out_channels (int): The output channels of this Module.
kernel_sizes (tuple[int]): Sequential of kernel sizes of pooling
layers. Default: (5, 9, 13).
conv_cfg (dict): Config dict for convolution layer. Default: None,
which means using conv2d.
norm_cfg (dict): Config dict for normalization layer.
Default: dict(type='BN').
act_cfg (dict): Config dict for activation layer.
Default: dict(type='Swish').
init_cfg (dict or list[dict], optional): Initialization config dict.
Default: None.
"""
def __init__(self,
in_channels,
out_channels,
kernel_sizes=(5, 9, 13),
conv_cfg=None,
norm_cfg=dict(type='BN', momentum=0.03, eps=0.001),
act_cfg=dict(type='Swish')):
super().__init__()
mid_channels = in_channels // 2
self.conv1 = ConvModule(
in_channels,
mid_channels,
1,
stride=1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg)
self.poolings = nn.ModuleList([
nn.MaxPool2d(kernel_size=ks, stride=1, padding=ks // 2)
for ks in kernel_sizes
])
conv2_channels = mid_channels * (len(kernel_sizes) + 1)
self.conv2 = ConvModule(
conv2_channels,
out_channels,
1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg)
def forward(self, x):
x = self.conv1(x)
x = torch.cat([x] + [pooling(x) for pooling in self.poolings], dim=1)
x = self.conv2(x)
return x
class CSPDarknet(nn.Module):
"""CSP-Darknet backbone used in YOLOv5 and YOLOX.
Args:
arch (str): Architecture of CSP-Darknet, from {P5, P6}.
Default: P5.
deepen_factor (float): Depth multiplier, multiply number of
blocks in CSP layer by this amount. Default: 1.0.
widen_factor (float): Width multiplier, multiply number of
channels in each layer by this amount. Default: 1.0.
out_indices (Sequence[int]): Output from which stages.
Default: (2, 3, 4).
frozen_stages (int): Stages to be frozen (stop grad and set eval
mode). -1 means not freezing any parameters. Default: -1.
use_depthwise (bool): Whether to use depthwise separable convolution.
Default: False.
arch_ovewrite(list): Overwrite default arch settings. Default: None.
spp_kernal_sizes: (tuple[int]): Sequential of kernel sizes of SPP
layers. Default: (5, 9, 13).
conv_cfg (dict): Config dict for convolution layer. Default: None.
norm_cfg (dict): Dictionary to construct and config norm layer.
Default: dict(type='BN', requires_grad=True).
act_cfg (dict): Config dict for activation layer.
Default: dict(type='LeakyReLU', negative_slope=0.1).
norm_eval (bool): Whether to set norm layers to eval mode, namely,
freeze running stats (mean and var). Note: Effect on Batch Norm
and its variants only.
"""
# From left to right:
# in_channels, out_channels, num_blocks, add_identity, use_spp
arch_settings = {
'P5': [[64, 128, 3, True, False], [128, 256, 9, True, False],
[256, 512, 9, True, False], [512, 1024, 3, False, True]],
'P6': [[64, 128, 3, True, False], [128, 256, 9, True, False],
[256, 512, 9, True, False], [512, 768, 3, True, False],
[768, 1024, 3, False, True]]
}
def __init__(self,
arch='P5',
deepen_factor=1.0,
widen_factor=1.0,
out_indices=(2, 3, 4),
frozen_stages=-1,
use_depthwise=False,
arch_ovewrite=None,
spp_kernal_sizes=(5, 9, 13),
conv_cfg=None,
norm_cfg=dict(type='BN', momentum=0.03, eps=0.001),
act_cfg=dict(type='Swish'),
norm_eval=False):
super().__init__()
arch_setting = self.arch_settings[arch]
if arch_ovewrite:
arch_setting = arch_ovewrite
assert set(out_indices).issubset(
i for i in range(len(arch_setting) + 1))
if frozen_stages not in range(-1, len(arch_setting) + 1):
raise ValueError('frozen_stages must be in range(-1, '
'len(arch_setting) + 1). But received '
f'{frozen_stages}')
self.out_indices = out_indices
self.frozen_stages = frozen_stages
self.use_depthwise = use_depthwise
self.norm_eval = norm_eval
self.stem = Focus(
3,
int(arch_setting[0][0] * widen_factor),
kernel_size=3,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg)
self.layers = ['stem']
for i, (in_channels, out_channels, num_blocks, add_identity,
use_spp) in enumerate(arch_setting):
in_channels = int(in_channels * widen_factor)
out_channels = int(out_channels * widen_factor)
num_blocks = max(round(num_blocks * deepen_factor), 1)
stage = []
conv_layer = ConvModule(
in_channels,
out_channels,
3,
stride=2,
padding=1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg)
stage.append(conv_layer)
if use_spp:
spp = SPPBottleneck(
out_channels,
out_channels,
kernel_sizes=spp_kernal_sizes,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg)
stage.append(spp)
csp_layer = CSPLayer(
out_channels,
out_channels,
num_blocks=num_blocks,
add_identity=add_identity,
use_depthwise=use_depthwise,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg)
stage.append(csp_layer)
self.add_module(f'stage{i + 1}', nn.Sequential(*stage))
self.layers.append(f'stage{i + 1}')
def _freeze_stages(self):
if self.frozen_stages >= 0:
for i in range(self.frozen_stages + 1):
m = getattr(self, self.layers[i])
m.eval()
for param in m.parameters():
param.requires_grad = False
def train(self, mode=True):
super(CSPDarknet, self).train(mode)
self._freeze_stages()
if mode and self.norm_eval:
for m in self.modules():
if isinstance(m, _BatchNorm):
m.eval()
def forward(self, x):
outs = []
for i, layer_name in enumerate(self.layers):
layer = getattr(self, layer_name)
x = layer(x)
if i in self.out_indices:
outs.append(x)
return tuple(outs)

View File

@ -1,288 +0,0 @@
import torch
import torch.nn as nn
from mmcv.cnn import ConvModule
from mmcv.ops.nms import batched_nms
from mmcv.runner import force_fp32
from functools import partial
from .priors import MlvlPointGenerator
def multi_apply(func, *args, **kwargs):
"""Apply function to a list of arguments.
Note:
This function applies the ``func`` to multiple inputs and
map the multiple outputs of the ``func`` into different
list. Each list contains the same type of outputs corresponding
to different inputs.
Args:
func (Function): A function that will be applied to a list of
arguments
Returns:
tuple(list): A tuple containing multiple list, each list contains \
a kind of returned results by the function
"""
pfunc = partial(func, **kwargs) if kwargs else func
map_results = map(pfunc, *args)
return tuple(map(list, zip(*map_results)))
class YOLOXHead(nn.Module):
"""YOLOXHead head used in `YOLOX <https://arxiv.org/abs/2107.08430>`_.
Args:
num_classes (int): Number of categories excluding the background
category.
in_channels (int): Number of channels in the input feature map.
feat_channels (int): Number of hidden channels in stacking convs.
Default: 256
stacked_convs (int): Number of stacking convs of the head.
Default: 2.
strides (tuple): Downsample factor of each feature map.
use_depthwise (bool): Whether to depthwise separable convolution in
blocks. Default: False
dcn_on_last_conv (bool): If true, use dcn in the last layer of
towers. Default: False.
conv_bias (bool | str): If specified as `auto`, it will be decided by
the norm_cfg. Bias of conv will be set as True if `norm_cfg` is
None, otherwise False. Default: "auto".
conv_cfg (dict): Config dict for convolution layer. Default: None.
norm_cfg (dict): Config dict for normalization layer. Default: None.
act_cfg (dict): Config dict for activation layer. Default: None.
"""
def __init__(self,
num_classes,
in_channels,
feat_channels=256,
stacked_convs=2,
strides=[8, 16, 32],
use_depthwise=False,
dcn_on_last_conv=False,
conv_bias='auto',
conv_cfg=None,
norm_cfg=dict(type='BN', momentum=0.03, eps=0.001),
act_cfg=dict(type='Swish'),
nms_score_thr=0.3,
nms_iou_threshold=0.1):
super().__init__()
self.nms_config = dict(score_thr=nms_score_thr, nms=dict(iou_threshold=nms_iou_threshold))
self.num_classes = num_classes
self.cls_out_channels = num_classes
self.in_channels = in_channels
self.feat_channels = feat_channels
self.stacked_convs = stacked_convs
self.strides = strides
self.use_depthwise = use_depthwise
self.dcn_on_last_conv = dcn_on_last_conv
assert conv_bias == 'auto' or isinstance(conv_bias, bool)
self.conv_bias = conv_bias
self.use_sigmoid_cls = True
self.conv_cfg = conv_cfg
self.norm_cfg = norm_cfg
self.act_cfg = act_cfg
self.prior_generator = MlvlPointGenerator(strides, offset=0)
self.fp16_enabled = False
self._init_layers()
def _init_layers(self):
self.multi_level_cls_convs = nn.ModuleList()
self.multi_level_reg_convs = nn.ModuleList()
self.multi_level_conv_cls = nn.ModuleList()
self.multi_level_conv_reg = nn.ModuleList()
self.multi_level_conv_obj = nn.ModuleList()
for _ in self.strides:
self.multi_level_cls_convs.append(self._build_stacked_convs())
self.multi_level_reg_convs.append(self._build_stacked_convs())
conv_cls, conv_reg, conv_obj = self._build_predictor()
self.multi_level_conv_cls.append(conv_cls)
self.multi_level_conv_reg.append(conv_reg)
self.multi_level_conv_obj.append(conv_obj)
def _build_stacked_convs(self):
"""Initialize conv layers of a single level head."""
conv = ConvModule
stacked_convs = []
for i in range(self.stacked_convs):
chn = self.in_channels if i == 0 else self.feat_channels
if self.dcn_on_last_conv and i == self.stacked_convs - 1:
conv_cfg = dict(type='DCNv2')
else:
conv_cfg = self.conv_cfg
stacked_convs.append(
conv(
chn,
self.feat_channels,
3,
stride=1,
padding=1,
conv_cfg=conv_cfg,
norm_cfg=self.norm_cfg,
act_cfg=self.act_cfg,
bias=self.conv_bias))
return nn.Sequential(*stacked_convs)
def _build_predictor(self):
"""Initialize predictor layers of a single level head."""
conv_cls = nn.Conv2d(self.feat_channels, self.cls_out_channels, 1)
conv_reg = nn.Conv2d(self.feat_channels, 4, 1)
conv_obj = nn.Conv2d(self.feat_channels, 1, 1)
return conv_cls, conv_reg, conv_obj
def forward_single(self, x, cls_convs, reg_convs, conv_cls, conv_reg,
conv_obj):
"""Forward feature of a single scale level."""
cls_feat = cls_convs(x)
reg_feat = reg_convs(x)
cls_score = conv_cls(cls_feat)
bbox_pred = conv_reg(reg_feat)
objectness = conv_obj(reg_feat)
return cls_score, bbox_pred, objectness
def forward(self, feats):
"""Forward features from the upstream network.
Args:
feats (tuple[Tensor]): Features from the upstream network, each is
a 4D-tensor.
Returns:
tuple[Tensor]: A tuple of multi-level predication map, each is a
4D-tensor of shape (batch_size, 5+num_classes, height, width).
"""
return multi_apply(self.forward_single, feats,
self.multi_level_cls_convs,
self.multi_level_reg_convs,
self.multi_level_conv_cls,
self.multi_level_conv_reg,
self.multi_level_conv_obj)
@force_fp32(apply_to=('cls_scores', 'bbox_preds', 'objectnesses'))
def get_bboxes(self,
cls_scores,
bbox_preds,
objectnesses,
cfg=None):
"""Transform network outputs of a batch into bbox results.
Args:
cls_scores (list[Tensor]): Classification scores for all
scale levels, each is a 4D-tensor, has shape
(batch_size, num_priors * num_classes, H, W).
bbox_preds (list[Tensor]): Box energies / deltas for all
scale levels, each is a 4D-tensor, has shape
(batch_size, num_priors * 4, H, W).
objectnesses (list[Tensor], Optional): Score factor for
all scale level, each is a 4D-tensor, has shape
(batch_size, 1, H, W).
cfg (mmcv.Config, Optional): Test / postprocessing configuration,
if None, test_cfg would be used. Default None.
rescale (bool): If True, return boxes in original image space.
Default False.
with_nms (bool): If True, do nms before return boxes.
Default True.
Returns:
list[list[Tensor, Tensor]]: Each item in result_list is 2-tuple.
The first item is an (n, 5) tensor, where the first 4 columns
are bounding box positions (tl_x, tl_y, br_x, br_y) and the
5-th column is a score between 0 and 1. The second item is a
(n,) tensor where each item is the predicted class label of
the corresponding box.
"""
assert len(cls_scores) == len(bbox_preds) == len(objectnesses)
cfg = self.nms_config if cfg is None else cfg
featmap_sizes = [cls_score.shape[2:] for cls_score in cls_scores]
mlvl_priors = self.prior_generator.grid_priors(
featmap_sizes,
dtype=cls_scores[0].dtype,
device=cls_scores[0].device,
with_stride=True)
# flatten cls_scores, bbox_preds and objectness
flatten_cls_scores = [
cls_score.permute(0, 2, 3, 1).reshape(1, -1,
self.cls_out_channels)
for cls_score in cls_scores
]
flatten_bbox_preds = [
bbox_pred.permute(0, 2, 3, 1).reshape(1, -1, 4)
for bbox_pred in bbox_preds
]
flatten_objectness = [
objectness.permute(0, 2, 3, 1).reshape(1, -1)
for objectness in objectnesses
]
flatten_cls_scores = torch.cat(flatten_cls_scores, dim=1).sigmoid()
flatten_bbox_preds = torch.cat(flatten_bbox_preds, dim=1)
flatten_objectness = torch.cat(flatten_objectness, dim=1).sigmoid()
flatten_priors = torch.cat(mlvl_priors)
flatten_bboxes = self._bbox_decode(flatten_priors, flatten_bbox_preds)
result_list = []
for img_id in range(1):
cls_scores = flatten_cls_scores[img_id]
score_factor = flatten_objectness[img_id]
bboxes = flatten_bboxes[img_id]
result_list.append(
self._bboxes_nms(cls_scores, bboxes, score_factor, cfg))
return result_list
def _bbox_decode(self, priors, bbox_preds):
xys = (bbox_preds[..., :2] * priors[:, 2:]) + priors[:, :2]
whs = bbox_preds[..., 2:].exp() * priors[:, 2:]
tl_x = (xys[..., 0] - whs[..., 0] / 2)
tl_y = (xys[..., 1] - whs[..., 1] / 2)
br_x = (xys[..., 0] + whs[..., 0] / 2)
br_y = (xys[..., 1] + whs[..., 1] / 2)
decoded_bboxes = torch.stack([tl_x, tl_y, br_x, br_y], -1)
return decoded_bboxes
def _bboxes_nms(self, cls_scores, bboxes, score_factor, cfg):
max_scores, labels = torch.max(cls_scores, 1)
valid_mask = score_factor * max_scores >= cfg['score_thr']
bboxes = bboxes[valid_mask]
scores = max_scores[valid_mask] * score_factor[valid_mask]
labels = labels[valid_mask]
if labels.numel() == 0:
return bboxes, labels
else:
dets, keep = batched_nms(bboxes.float(), scores.float(), labels, cfg['nms'])
return dets, labels[keep]
def simple_test_bboxes(self, feats):
"""Test det bboxes without test-time augmentation, can be applied in
DenseHead except for ``RPNHead`` and its variants, e.g., ``GARPNHead``,
etc.
Args:
feats (tuple[torch.Tensor]): Multi-level features from the
upstream network, each is a 4D-tensor.
rescale (bool, optional): Whether to rescale the results.
Defaults to False.
Returns:
list[tuple[Tensor, Tensor]]: Each item in result_list is 2-tuple.
The first item is ``bboxes`` with shape (n, 5),
where 5 represent (tl_x, tl_y, br_x, br_y, score).
The shape of the second tensor in the tuple is ``labels``
with shape (n,)
"""
outs = self.forward(feats)
results_list = self.get_bboxes(*outs)
return results_list

View File

@ -1,75 +0,0 @@
import os
import shutil
import hashlib
import warnings
def sha256sum(filename):
h = hashlib.sha256()
b = bytearray(128*1024)
mv = memoryview(b)
with open(filename, 'rb', buffering=0) as f:
for n in iter(lambda : f.readinto(mv), 0):
h.update(mv[:n])
return h.hexdigest()
online_model_factory = {
'yolox-s-general-text-pretrain-20221226': {
'url': 'https://github.com/moewiee/satrn-model-factory/raw/main/62j266xm8r.pth',
'hash': '89bff792685af454d0cfea5d6d673be6914d614e4c2044e786da6eddf36f8b50'},
'yolox-s-checkbox-20220726': {
'url': 'https://github.com/moewiee/satrn-model-factory/raw/main/1647d7eys7.pth',
'hash': '7c1e188b7375dcf0b7b9d317675ebd92a86fdc29363558002249867249ee10f8'},
'yolox-s-idcard-5c-20221027': {
'url': 'https://github.com/moewiee/satrn-model-factory/raw/main/jr0egad3ix.pth',
'hash': '73a7772594c1f6d3f6d6a98b6d6e4097af5026864e3bd50531ad9e635ae795a7'},
'yolox-s-handwritten-text-line-20230228': {
'url': 'https://github.com/moewiee/satrn-model-factory/raw/main/rb07rtwmgi.pth',
'hash': 'a31d1bf8fc880479d2e11463dad0b4081952a13e553a02919109b634a1190ef1'}
}
__hub_available_versions__ = online_model_factory.keys()
def _get_from_hub(file_path, version, version_url):
os.system(f'wget -O {file_path} {version_url}')
assert os.path.exists(file_path), \
'wget failed while trying to retrieve from hub.'
downloaded_hash = sha256sum(file_path)
if downloaded_hash != online_model_factory[version]['hash']:
os.remove(file_path)
raise ValueError('sha256 hash doesnt match for version retrieved from hub.')
def _get(version):
use_online = version in __hub_available_versions__
if not use_online and not os.path.exists(version):
raise ValueError(f'Model version {version} not found online and not found local.')
hub_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), 'hub')
if not os.path.exists(hub_path):
os.makedirs(hub_path)
if use_online:
version_url = online_model_factory[version]['url']
file_path = os.path.join(hub_path, os.path.basename(version_url))
else:
file_path = os.path.join(hub_path, os.path.basename(version))
if not os.path.exists(file_path):
if use_online:
_get_from_hub(file_path, version, version_url)
else:
shutil.copy2(version, file_path)
else:
if use_online:
downloaded_hash = sha256sum(file_path)
if downloaded_hash != online_model_factory[version]['hash']:
os.remove(file_path)
warnings.warn('existing hub version sha256 hash doesnt match, now re-download from hub.')
_get_from_hub(file_path, version, version_url)
else:
if sha256sum(file_path) != sha256sum(version):
os.remove(file_path)
warnings.warn('existing local version sha256 hash doesnt match, now replace with new local version.')
shutil.copy2(version, file_path)
return file_path

View File

@ -1,151 +0,0 @@
import torch
import torch.nn as nn
import numpy as np
from .backbone import CSPDarknet
from .neck import YOLOXPAFPN
from .bbox_head import YOLOXHead
from .transform import DetectorDataPipeline, AutoRotateDetectorDataPipeline
from .factory import _get as get_version
def bbox2result(bboxes, labels, num_classes):
"""Convert detection results to a list of numpy arrays.
Args:
bboxes (torch.Tensor | np.ndarray): shape (n, 5)
labels (torch.Tensor | np.ndarray): shape (n, )
num_classes (int): class number, including background class
Returns:
list(ndarray): bbox results of each class
"""
if bboxes.shape[0] == 0:
return [np.zeros((0, 5), dtype=np.float32) for i in range(num_classes)]
else:
if isinstance(bboxes, torch.Tensor):
bboxes = bboxes.detach().cpu().numpy()
labels = labels.detach().cpu().numpy()
return [bboxes[labels == i, :] for i in range(num_classes)]
def normalize_bbox(bboxes, scale):
for i in range(len(bboxes)):
bboxes[i][:,:4] /= scale
return bboxes
class SingleStageDetector(nn.Module):
def __init__(self,
version,
device):
super(SingleStageDetector, self).__init__()
assert 'cpu' in device or 'cuda' in device
checkpoint = get_version(version)
pt = torch.load(checkpoint, 'cpu')
self.pipeline = DetectorDataPipeline(**pt['pipeline_args'], device=device)
self.backbone = CSPDarknet(**pt['backbone_args'])
self.neck = YOLOXPAFPN(**pt['neck_args'])
self.bbox_head = YOLOXHead(**pt['bbox_head_args'])
self.load_state_dict(pt['state_dict'], strict=True)
self.eval()
for param in self.parameters():
param.requires_grad = False
self = self.to(device=device)
print(f'Text detection load from version {version}.')
def extract_feat(self, img):
"""Directly extract features from the backbone + neck."""
x = self.backbone(img)
x = self.neck(x)
return x
def forward(self, img):
"""Test function without test-time augmentation.
Args:
img (np.ndarray): Images with shape (H, W, C) or
img (str): Path to image.
Returns:
list[list[np.ndarray]]: BBox results of each image and classes.
The list corresponds to each class.
"""
img, origin_shape, new_shape = self.pipeline(img)
scale = min(new_shape / origin_shape)
feat = self.extract_feat(img)
results_list = self.bbox_head.simple_test_bboxes(feat)
bbox_results = [
bbox2result(det_bboxes, det_labels, self.bbox_head.num_classes)
for det_bboxes, det_labels in results_list
][0]
bbox_results = normalize_bbox(bbox_results, scale)
return bbox_results
class AutoRotateDetector(nn.Module):
def __init__(self,
version,
device):
super(AutoRotateDetector, self).__init__()
assert 'cpu' in device or 'cuda' in device
checkpoint = get_version(version)
pt = torch.load(checkpoint, 'cpu')
self.pipeline = AutoRotateDetectorDataPipeline(**pt['pipeline_args'], device=device)
self.backbone = CSPDarknet(**pt['backbone_args'])
self.neck = YOLOXPAFPN(**pt['neck_args'])
self.bbox_head = YOLOXHead(**pt['bbox_head_args'], nms_score_thr=0.8)
self.load_state_dict(pt['state_dict'], strict=True)
self.eval()
for param in self.parameters():
param.requires_grad = False
self = self.to(device=device)
print(f'Auto rotate detector load from version {version}.')
def extract_feat(self, img):
"""Directly extract features from the backbone + neck."""
x = self.backbone(img)
x = self.neck(x)
return x
def forward(self, img):
"""Test function without test-time augmentation.
Args:
img (np.ndarray): Images with shape (H, W, C) or
img (str): Path to image.
Returns:
np.ndarray: Straight rotated image.
"""
imgs, imgs_np = self.pipeline(img)
maxCount = -1
maxCountRot = None
for idx, img in enumerate(imgs):
currentCount = 0
feat = self.extract_feat(img)
results_list = self.bbox_head.simple_test_bboxes(feat)
bbox_results = [
bbox2result(det_bboxes, det_labels, self.bbox_head.num_classes)
for det_bboxes, det_labels in results_list
][0]
for class_result in bbox_results:
currentCount += len(class_result)
if currentCount > maxCount:
maxCount = currentCount
maxCountRot = idx
return imgs_np[maxCountRot]

View File

@ -1,140 +0,0 @@
import torch
import torch.nn as nn
from mmcv.cnn import ConvModule
from .backbone import CSPLayer
class YOLOXPAFPN(nn.Module):
"""Path Aggregation Network used in YOLOX.
Args:
in_channels (List[int]): Number of input channels per scale.
out_channels (int): Number of output channels (used at each scale)
num_csp_blocks (int): Number of bottlenecks in CSPLayer. Default: 3
use_depthwise (bool): Whether to depthwise separable convolution in
blocks. Default: False
upsample_cfg (dict): Config dict for interpolate layer.
Default: `dict(scale_factor=2, mode='nearest')`
conv_cfg (dict, optional): Config dict for convolution layer.
Default: None, which means using conv2d.
norm_cfg (dict): Config dict for normalization layer.
Default: dict(type='BN')
act_cfg (dict): Config dict for activation layer.
Default: dict(type='Swish')
init_cfg (dict or list[dict], optional): Initialization config dict.
Default: None.
"""
def __init__(self,
in_channels,
out_channels,
num_csp_blocks=3,
use_depthwise=False,
upsample_cfg=dict(scale_factor=2, mode='nearest'),
conv_cfg=None,
norm_cfg=dict(type='BN', momentum=0.03, eps=0.001),
act_cfg=dict(type='Swish')):
super(YOLOXPAFPN, self).__init__()
self.in_channels = in_channels
self.out_channels = out_channels
# build top-down blocks
self.upsample = nn.Upsample(**upsample_cfg)
self.reduce_layers = nn.ModuleList()
self.top_down_blocks = nn.ModuleList()
for idx in range(len(in_channels) - 1, 0, -1):
self.reduce_layers.append(
ConvModule(
in_channels[idx],
in_channels[idx - 1],
1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg))
self.top_down_blocks.append(
CSPLayer(
in_channels[idx - 1] * 2,
in_channels[idx - 1],
num_blocks=num_csp_blocks,
add_identity=False,
use_depthwise=use_depthwise,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg))
# build bottom-up blocks
self.downsamples = nn.ModuleList()
self.bottom_up_blocks = nn.ModuleList()
for idx in range(len(in_channels) - 1):
self.downsamples.append(
ConvModule(
in_channels[idx],
in_channels[idx],
3,
stride=2,
padding=1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg))
self.bottom_up_blocks.append(
CSPLayer(
in_channels[idx] * 2,
in_channels[idx + 1],
num_blocks=num_csp_blocks,
add_identity=False,
use_depthwise=use_depthwise,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg))
self.out_convs = nn.ModuleList()
for i in range(len(in_channels)):
self.out_convs.append(
ConvModule(
in_channels[i],
out_channels,
1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg))
def forward(self, inputs):
"""
Args:
inputs (tuple[Tensor]): input features.
Returns:
tuple[Tensor]: YOLOXPAFPN features.
"""
assert len(inputs) == len(self.in_channels)
# top-down path
inner_outs = [inputs[-1]]
for idx in range(len(self.in_channels) - 1, 0, -1):
feat_heigh = inner_outs[0]
feat_low = inputs[idx - 1]
feat_heigh = self.reduce_layers[len(self.in_channels) - 1 - idx](
feat_heigh)
inner_outs[0] = feat_heigh
upsample_feat = self.upsample(feat_heigh)
inner_out = self.top_down_blocks[len(self.in_channels) - 1 - idx](
torch.cat([upsample_feat, feat_low], 1))
inner_outs.insert(0, inner_out)
# bottom-up path
outs = [inner_outs[0]]
for idx in range(len(self.in_channels) - 1):
feat_low = outs[-1]
feat_height = inner_outs[idx + 1]
downsample_feat = self.downsamples[idx](feat_low)
out = self.bottom_up_blocks[idx](
torch.cat([downsample_feat, feat_height], 1))
outs.append(out)
# out convs
for idx, conv in enumerate(self.out_convs):
outs[idx] = conv(outs[idx])
return tuple(outs)

View File

@ -1,225 +0,0 @@
import torch
import numpy as np
from torch.nn.modules.utils import _pair
class MlvlPointGenerator:
"""Standard points generator for multi-level (Mlvl) feature maps in 2D
points-based detectors.
Args:
strides (list[int] | list[tuple[int, int]]): Strides of anchors
in multiple feature levels in order (w, h).
offset (float): The offset of points, the value is normalized with
corresponding stride. Defaults to 0.5.
"""
def __init__(self, strides, offset=0.5):
self.strides = [_pair(stride) for stride in strides]
self.offset = offset
@property
def num_levels(self):
"""int: number of feature levels that the generator will be applied"""
return len(self.strides)
@property
def num_base_priors(self):
"""list[int]: The number of priors (points) at a point
on the feature grid"""
return [1 for _ in range(len(self.strides))]
def _meshgrid(self, x, y, row_major=True):
yy, xx = torch.meshgrid(y, x)
if row_major:
# warning .flatten() would cause error in ONNX exporting
# have to use reshape here
return xx.flatten(), yy.flatten()
else:
return yy.flatten(), xx.flatten()
def grid_priors(self,
featmap_sizes,
dtype=torch.float32,
device='cuda',
with_stride=False):
"""Generate grid points of multiple feature levels.
Args:
featmap_sizes (list[tuple]): List of feature map sizes in
multiple feature levels, each size arrange as
as (h, w).
dtype (:obj:`dtype`): Dtype of priors. Default: torch.float32.
device (str): The device where the anchors will be put on.
with_stride (bool): Whether to concatenate the stride to
the last dimension of points.
Return:
list[torch.Tensor]: Points of multiple feature levels.
The sizes of each tensor should be (N, 2) when with stride is
``False``, where N = width * height, width and height
are the sizes of the corresponding feature level,
and the last dimension 2 represent (coord_x, coord_y),
otherwise the shape should be (N, 4),
and the last dimension 4 represent
(coord_x, coord_y, stride_w, stride_h).
"""
assert self.num_levels == len(featmap_sizes)
multi_level_priors = []
for i in range(self.num_levels):
priors = self.single_level_grid_priors(
featmap_sizes[i],
level_idx=i,
dtype=dtype,
device=device,
with_stride=with_stride)
multi_level_priors.append(priors)
return multi_level_priors
def single_level_grid_priors(self,
featmap_size,
level_idx,
dtype=torch.float32,
device='cuda',
with_stride=False):
"""Generate grid Points of a single level.
Note:
This function is usually called by method ``self.grid_priors``.
Args:
featmap_size (tuple[int]): Size of the feature maps, arrange as
(h, w).
level_idx (int): The index of corresponding feature map level.
dtype (:obj:`dtype`): Dtype of priors. Default: torch.float32.
device (str, optional): The device the tensor will be put on.
Defaults to 'cuda'.
with_stride (bool): Concatenate the stride to the last dimension
of points.
Return:
Tensor: Points of single feature levels.
The shape of tensor should be (N, 2) when with stride is
``False``, where N = width * height, width and height
are the sizes of the corresponding feature level,
and the last dimension 2 represent (coord_x, coord_y),
otherwise the shape should be (N, 4),
and the last dimension 4 represent
(coord_x, coord_y, stride_w, stride_h).
"""
feat_h, feat_w = featmap_size
stride_w, stride_h = self.strides[level_idx]
shift_x = (torch.arange(0, feat_w, device=device) +
self.offset) * stride_w
# keep featmap_size as Tensor instead of int, so that we
# can convert to ONNX correctly
shift_x = shift_x.to(dtype)
shift_y = (torch.arange(0, feat_h, device=device) +
self.offset) * stride_h
# keep featmap_size as Tensor instead of int, so that we
# can convert to ONNX correctly
shift_y = shift_y.to(dtype)
shift_xx, shift_yy = self._meshgrid(shift_x, shift_y)
if not with_stride:
shifts = torch.stack([shift_xx, shift_yy], dim=-1)
else:
# use `shape[0]` instead of `len(shift_xx)` for ONNX export
stride_w = shift_xx.new_full((shift_xx.shape[0], ),
stride_w).to(dtype)
stride_h = shift_xx.new_full((shift_yy.shape[0], ),
stride_h).to(dtype)
shifts = torch.stack([shift_xx, shift_yy, stride_w, stride_h],
dim=-1)
all_points = shifts.to(device)
return all_points
def valid_flags(self, featmap_sizes, pad_shape, device='cuda'):
"""Generate valid flags of points of multiple feature levels.
Args:
featmap_sizes (list(tuple)): List of feature map sizes in
multiple feature levels, each size arrange as
as (h, w).
pad_shape (tuple(int)): The padded shape of the image,
arrange as (h, w).
device (str): The device where the anchors will be put on.
Return:
list(torch.Tensor): Valid flags of points of multiple levels.
"""
assert self.num_levels == len(featmap_sizes)
multi_level_flags = []
for i in range(self.num_levels):
point_stride = self.strides[i]
feat_h, feat_w = featmap_sizes[i]
h, w = pad_shape[:2]
valid_feat_h = min(int(np.ceil(h / point_stride[1])), feat_h)
valid_feat_w = min(int(np.ceil(w / point_stride[0])), feat_w)
flags = self.single_level_valid_flags((feat_h, feat_w),
(valid_feat_h, valid_feat_w),
device=device)
multi_level_flags.append(flags)
return multi_level_flags
def single_level_valid_flags(self,
featmap_size,
valid_size,
device='cuda'):
"""Generate the valid flags of points of a single feature map.
Args:
featmap_size (tuple[int]): The size of feature maps, arrange as
as (h, w).
valid_size (tuple[int]): The valid size of the feature maps.
The size arrange as as (h, w).
device (str, optional): The device where the flags will be put on.
Defaults to 'cuda'.
Returns:
torch.Tensor: The valid flags of each points in a single level \
feature map.
"""
feat_h, feat_w = featmap_size
valid_h, valid_w = valid_size
assert valid_h <= feat_h and valid_w <= feat_w
valid_x = torch.zeros(feat_w, dtype=torch.bool, device=device)
valid_y = torch.zeros(feat_h, dtype=torch.bool, device=device)
valid_x[:valid_w] = 1
valid_y[:valid_h] = 1
valid_xx, valid_yy = self._meshgrid(valid_x, valid_y)
valid = valid_xx & valid_yy
return valid
def sparse_priors(self,
prior_idxs,
featmap_size,
level_idx,
dtype=torch.float32,
device='cuda'):
"""Generate sparse points according to the ``prior_idxs``.
Args:
prior_idxs (Tensor): The index of corresponding anchors
in the feature map.
featmap_size (tuple[int]): feature map size arrange as (w, h).
level_idx (int): The level index of corresponding feature
map.
dtype (obj:`torch.dtype`): Date type of points. Defaults to
``torch.float32``.
device (obj:`torch.device`): The device where the points is
located.
Returns:
Tensor: Anchor with shape (N, 2), N should be equal to
the length of ``prior_idxs``. And last dimension
2 represent (coord_x, coord_y).
"""
height, width = featmap_size
x = (prior_idxs % width + self.offset) * self.strides[level_idx][0]
y = ((prior_idxs // width) % height +
self.offset) * self.strides[level_idx][1]
prioris = torch.stack([x, y], 1).to(dtype)
prioris = prioris.to(device)
return prioris

View File

@ -1,81 +0,0 @@
import mmcv
import numpy as np
import cv2
import torch
class DetectorDataPipeline:
def __init__(self,
img_scale,
device):
self.scale = img_scale
self.device = device
def load(self, img):
if isinstance(img, str):
return cv2.imread(img)
elif isinstance(img, np.ndarray):
return img
else:
raise ValueError(f'img input must be a str/np.ndarray, got {type(img)}')
def resize(self, img):
origin_shape = img.shape[:2]
img = mmcv.imrescale(img,
self.scale,
return_scale=False,
interpolation='bilinear',
backend='cv2')
return img, origin_shape, np.array(self.scale)
def pad(self, img):
if self.scale is not None:
width = max(self.scale[1] - img.shape[1], 0)
height = max(self.scale[0] - img.shape[0], 0)
padding = (0, 0, width, height)
img = cv2.copyMakeBorder(img,
padding[1],
padding[3],
padding[0],
padding[2],
cv2.BORDER_CONSTANT,
value=(114, 114, 114,))
return img
def to_tensor(self, img):
img = torch.from_numpy(img.astype(np.float32)).permute(2,0,1).unsqueeze(0)
img = img.to(device=self.device)
return img
def __call__(self, img):
img = self.load(img)
img, origin_shape, new_shape = self.resize(img)
img = self.pad(img)
img = self.to_tensor(img)
return img, origin_shape, new_shape
class AutoRotateDetectorDataPipeline(DetectorDataPipeline):
def __call__(self, img):
img = self.load(img)
imgs = []
imgs_np = []
for flag in (None, cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_180, cv2.ROTATE_90_COUNTERCLOCKWISE):
img_t = img if flag is None else cv2.rotate(img, flag)
imgs_np.append(img_t)
img_t, _, _ = self.resize(img_t)
img_t = self.pad(img_t)
img_t = self.to_tensor(img_t)
imgs.append(img_t)
return imgs, imgs_np

View File

@ -1,186 +0,0 @@
import os
import os.path as osp
import shutil
import sys
import warnings
from setuptools import find_packages, setup
def readme():
with open('README.md', encoding='utf-8') as f:
content = f.read()
return content
version_file = 'sdsvtd/version.py'
is_windows = sys.platform == 'win32'
def add_mim_extention():
"""Add extra files that are required to support MIM into the package.
These files will be added by creating a symlink to the originals if the
package is installed in `editable` mode (e.g. pip install -e .), or by
copying from the originals otherwise.
"""
# parse installment mode
if 'develop' in sys.argv:
# installed by `pip install -e .`
mode = 'symlink'
elif 'sdist' in sys.argv or 'bdist_wheel' in sys.argv:
# installed by `pip install .`
# or create source distribution by `python setup.py sdist`
mode = 'copy'
else:
return
filenames = ['tools', 'configs', 'model-index.yml']
repo_path = osp.dirname(__file__)
mim_path = osp.join(repo_path, 'mmocr', '.mim')
os.makedirs(mim_path, exist_ok=True)
for filename in filenames:
if osp.exists(filename):
src_path = osp.join(repo_path, filename)
tar_path = osp.join(mim_path, filename)
if osp.isfile(tar_path) or osp.islink(tar_path):
os.remove(tar_path)
elif osp.isdir(tar_path):
shutil.rmtree(tar_path)
if mode == 'symlink':
src_relpath = osp.relpath(src_path, osp.dirname(tar_path))
try:
os.symlink(src_relpath, tar_path)
except OSError:
# Creating a symbolic link on windows may raise an
# `OSError: [WinError 1314]` due to privilege. If
# the error happens, the src file will be copied
mode = 'copy'
warnings.warn(
f'Failed to create a symbolic link for {src_relpath}, '
f'and it will be copied to {tar_path}')
else:
continue
if mode == 'copy':
if osp.isfile(src_path):
shutil.copyfile(src_path, tar_path)
elif osp.isdir(src_path):
shutil.copytree(src_path, tar_path)
else:
warnings.warn(f'Cannot copy file {src_path}.')
else:
raise ValueError(f'Invalid mode {mode}')
def get_version():
with open(version_file, 'r') as f:
exec(compile(f.read(), version_file, 'exec'))
import sys
# return short version for sdist
if 'sdist' in sys.argv or 'bdist_wheel' in sys.argv:
return locals()['short_version']
else:
return locals()['__version__']
def parse_requirements(fname='requirements.txt', with_version=True):
"""Parse the package dependencies listed in a requirements file but strip
specific version information.
Args:
fname (str): Path to requirements file.
with_version (bool, default=False): If True, include version specs.
Returns:
info (list[str]): List of requirements items.
CommandLine:
python -c "import setup; print(setup.parse_requirements())"
"""
import re
import sys
from os.path import exists
require_fpath = fname
def parse_line(line):
"""Parse information from a line in a requirements text file."""
if line.startswith('-r '):
# Allow specifying requirements in other files
target = line.split(' ')[1]
for info in parse_require_file(target):
yield info
else:
info = {'line': line}
if line.startswith('-e '):
info['package'] = line.split('#egg=')[1]
else:
# Remove versioning from the package
pat = '(' + '|'.join(['>=', '==', '>']) + ')'
parts = re.split(pat, line, maxsplit=1)
parts = [p.strip() for p in parts]
info['package'] = parts[0]
if len(parts) > 1:
op, rest = parts[1:]
if ';' in rest:
# Handle platform specific dependencies
# http://setuptools.readthedocs.io/en/latest/setuptools.html#declaring-platform-specific-dependencies
version, platform_deps = map(str.strip,
rest.split(';'))
info['platform_deps'] = platform_deps
else:
version = rest # NOQA
info['version'] = (op, version)
yield info
def parse_require_file(fpath):
with open(fpath, 'r') as f:
for line in f.readlines():
line = line.strip()
if line and not line.startswith('#'):
for info in parse_line(line):
yield info
def gen_packages_items():
if exists(require_fpath):
for info in parse_require_file(require_fpath):
parts = [info['package']]
if with_version and 'version' in info:
parts.extend(info['version'])
if not sys.version.startswith('3.4'):
# apparently package_deps are broken in 3.4
platform_deps = info.get('platform_deps')
if platform_deps is not None:
parts.append(';' + platform_deps)
item = ''.join(parts)
yield item
packages = list(gen_packages_items())
return packages
if __name__ == '__main__':
setup(
name='sdsvtd',
version=get_version(),
description='SDSV OCR Team Text Detection',
long_description=readme(),
long_description_content_type='text/markdown',
packages=find_packages(exclude=('configs', 'tools', 'demo')),
include_package_data=True,
classifiers=[
'Development Status :: 4 - Beta',
'License :: OSI Approved :: Apache Software License',
'Operating System :: OS Independent',
'Programming Language :: Python :: 3',
'Programming Language :: Python :: 3.6',
'Programming Language :: Python :: 3.7',
'Programming Language :: Python :: 3.8',
'Programming Language :: Python :: 3.9',
],
license='Apache License 2.0',
install_requires=parse_requirements('requirements.txt'),
zip_safe=False)

View File

@ -1,6 +0,0 @@
# Builds
*.egg-info
__pycache__
# Checkpoint
hub

View File

@ -1,674 +0,0 @@
GNU GENERAL PUBLIC LICENSE
Version 3, 29 June 2007
Copyright (C) 2007 Free Software Foundation, Inc. <http://fsf.org/>
Everyone is permitted to copy and distribute verbatim copies
of this license document, but changing it is not allowed.
Preamble
The GNU General Public License is a free, copyleft license for
software and other kinds of works.
The licenses for most software and other practical works are designed
to take away your freedom to share and change the works. By contrast,
the GNU General Public License is intended to guarantee your freedom to
share and change all versions of a program--to make sure it remains free
software for all its users. We, the Free Software Foundation, use the
GNU General Public License for most of our software; it applies also to
any other work released this way by its authors. You can apply it to
your programs, too.
When we speak of free software, we are referring to freedom, not
price. Our General Public Licenses are designed to make sure that you
have the freedom to distribute copies of free software (and charge for
them if you wish), that you receive source code or can get it if you
want it, that you can change the software or use pieces of it in new
free programs, and that you know you can do these things.
To protect your rights, we need to prevent others from denying you
these rights or asking you to surrender the rights. Therefore, you have
certain responsibilities if you distribute copies of the software, or if
you modify it: responsibilities to respect the freedom of others.
For example, if you distribute copies of such a program, whether
gratis or for a fee, you must pass on to the recipients the same
freedoms that you received. You must make sure that they, too, receive
or can get the source code. And you must show them these terms so they
know their rights.
Developers that use the GNU GPL protect your rights with two steps:
(1) assert copyright on the software, and (2) offer you this License
giving you legal permission to copy, distribute and/or modify it.
For the developers' and authors' protection, the GPL clearly explains
that there is no warranty for this free software. For both users' and
authors' sake, the GPL requires that modified versions be marked as
changed, so that their problems will not be attributed erroneously to
authors of previous versions.
Some devices are designed to deny users access to install or run
modified versions of the software inside them, although the manufacturer
can do so. This is fundamentally incompatible with the aim of
protecting users' freedom to change the software. The systematic
pattern of such abuse occurs in the area of products for individuals to
use, which is precisely where it is most unacceptable. Therefore, we
have designed this version of the GPL to prohibit the practice for those
products. If such problems arise substantially in other domains, we
stand ready to extend this provision to those domains in future versions
of the GPL, as needed to protect the freedom of users.
Finally, every program is threatened constantly by software patents.
States should not allow patents to restrict development and use of
software on general-purpose computers, but in those that do, we wish to
avoid the special danger that patents applied to a free program could
make it effectively proprietary. To prevent this, the GPL assures that
patents cannot be used to render the program non-free.
The precise terms and conditions for copying, distribution and
modification follow.
TERMS AND CONDITIONS
0. Definitions.
"This License" refers to version 3 of the GNU General Public License.
"Copyright" also means copyright-like laws that apply to other kinds of
works, such as semiconductor masks.
"The Program" refers to any copyrightable work licensed under this
License. Each licensee is addressed as "you". "Licensees" and
"recipients" may be individuals or organizations.
To "modify" a work means to copy from or adapt all or part of the work
in a fashion requiring copyright permission, other than the making of an
exact copy. The resulting work is called a "modified version" of the
earlier work or a work "based on" the earlier work.
A "covered work" means either the unmodified Program or a work based
on the Program.
To "propagate" a work means to do anything with it that, without
permission, would make you directly or secondarily liable for
infringement under applicable copyright law, except executing it on a
computer or modifying a private copy. Propagation includes copying,
distribution (with or without modification), making available to the
public, and in some countries other activities as well.
To "convey" a work means any kind of propagation that enables other
parties to make or receive copies. Mere interaction with a user through
a computer network, with no transfer of a copy, is not conveying.
An interactive user interface displays "Appropriate Legal Notices"
to the extent that it includes a convenient and prominently visible
feature that (1) displays an appropriate copyright notice, and (2)
tells the user that there is no warranty for the work (except to the
extent that warranties are provided), that licensees may convey the
work under this License, and how to view a copy of this License. If
the interface presents a list of user commands or options, such as a
menu, a prominent item in the list meets this criterion.
1. Source Code.
The "source code" for a work means the preferred form of the work
for making modifications to it. "Object code" means any non-source
form of a work.
A "Standard Interface" means an interface that either is an official
standard defined by a recognized standards body, or, in the case of
interfaces specified for a particular programming language, one that
is widely used among developers working in that language.
The "System Libraries" of an executable work include anything, other
than the work as a whole, that (a) is included in the normal form of
packaging a Major Component, but which is not part of that Major
Component, and (b) serves only to enable use of the work with that
Major Component, or to implement a Standard Interface for which an
implementation is available to the public in source code form. A
"Major Component", in this context, means a major essential component
(kernel, window system, and so on) of the specific operating system
(if any) on which the executable work runs, or a compiler used to
produce the work, or an object code interpreter used to run it.
The "Corresponding Source" for a work in object code form means all
the source code needed to generate, install, and (for an executable
work) run the object code and to modify the work, including scripts to
control those activities. However, it does not include the work's
System Libraries, or general-purpose tools or generally available free
programs which are used unmodified in performing those activities but
which are not part of the work. For example, Corresponding Source
includes interface definition files associated with source files for
the work, and the source code for shared libraries and dynamically
linked subprograms that the work is specifically designed to require,
such as by intimate data communication or control flow between those
subprograms and other parts of the work.
The Corresponding Source need not include anything that users
can regenerate automatically from other parts of the Corresponding
Source.
The Corresponding Source for a work in source code form is that
same work.
2. Basic Permissions.
All rights granted under this License are granted for the term of
copyright on the Program, and are irrevocable provided the stated
conditions are met. This License explicitly affirms your unlimited
permission to run the unmodified Program. The output from running a
covered work is covered by this License only if the output, given its
content, constitutes a covered work. This License acknowledges your
rights of fair use or other equivalent, as provided by copyright law.
You may make, run and propagate covered works that you do not
convey, without conditions so long as your license otherwise remains
in force. You may convey covered works to others for the sole purpose
of having them make modifications exclusively for you, or provide you
with facilities for running those works, provided that you comply with
the terms of this License in conveying all material for which you do
not control copyright. Those thus making or running the covered works
for you must do so exclusively on your behalf, under your direction
and control, on terms that prohibit them from making any copies of
your copyrighted material outside their relationship with you.
Conveying under any other circumstances is permitted solely under
the conditions stated below. Sublicensing is not allowed; section 10
makes it unnecessary.
3. Protecting Users' Legal Rights From Anti-Circumvention Law.
No covered work shall be deemed part of an effective technological
measure under any applicable law fulfilling obligations under article
11 of the WIPO copyright treaty adopted on 20 December 1996, or
similar laws prohibiting or restricting circumvention of such
measures.
When you convey a covered work, you waive any legal power to forbid
circumvention of technological measures to the extent such circumvention
is effected by exercising rights under this License with respect to
the covered work, and you disclaim any intention to limit operation or
modification of the work as a means of enforcing, against the work's
users, your or third parties' legal rights to forbid circumvention of
technological measures.
4. Conveying Verbatim Copies.
You may convey verbatim copies of the Program's source code as you
receive it, in any medium, provided that you conspicuously and
appropriately publish on each copy an appropriate copyright notice;
keep intact all notices stating that this License and any
non-permissive terms added in accord with section 7 apply to the code;
keep intact all notices of the absence of any warranty; and give all
recipients a copy of this License along with the Program.
You may charge any price or no price for each copy that you convey,
and you may offer support or warranty protection for a fee.
5. Conveying Modified Source Versions.
You may convey a work based on the Program, or the modifications to
produce it from the Program, in the form of source code under the
terms of section 4, provided that you also meet all of these conditions:
a) The work must carry prominent notices stating that you modified
it, and giving a relevant date.
b) The work must carry prominent notices stating that it is
released under this License and any conditions added under section
7. This requirement modifies the requirement in section 4 to
"keep intact all notices".
c) You must license the entire work, as a whole, under this
License to anyone who comes into possession of a copy. This
License will therefore apply, along with any applicable section 7
additional terms, to the whole of the work, and all its parts,
regardless of how they are packaged. This License gives no
permission to license the work in any other way, but it does not
invalidate such permission if you have separately received it.
d) If the work has interactive user interfaces, each must display
Appropriate Legal Notices; however, if the Program has interactive
interfaces that do not display Appropriate Legal Notices, your
work need not make them do so.
A compilation of a covered work with other separate and independent
works, which are not by their nature extensions of the covered work,
and which are not combined with it such as to form a larger program,
in or on a volume of a storage or distribution medium, is called an
"aggregate" if the compilation and its resulting copyright are not
used to limit the access or legal rights of the compilation's users
beyond what the individual works permit. Inclusion of a covered work
in an aggregate does not cause this License to apply to the other
parts of the aggregate.
6. Conveying Non-Source Forms.
You may convey a covered work in object code form under the terms
of sections 4 and 5, provided that you also convey the
machine-readable Corresponding Source under the terms of this License,
in one of these ways:
a) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by the
Corresponding Source fixed on a durable physical medium
customarily used for software interchange.
b) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by a
written offer, valid for at least three years and valid for as
long as you offer spare parts or customer support for that product
model, to give anyone who possesses the object code either (1) a
copy of the Corresponding Source for all the software in the
product that is covered by this License, on a durable physical
medium customarily used for software interchange, for a price no
more than your reasonable cost of physically performing this
conveying of source, or (2) access to copy the
Corresponding Source from a network server at no charge.
c) Convey individual copies of the object code with a copy of the
written offer to provide the Corresponding Source. This
alternative is allowed only occasionally and noncommercially, and
only if you received the object code with such an offer, in accord
with subsection 6b.
d) Convey the object code by offering access from a designated
place (gratis or for a charge), and offer equivalent access to the
Corresponding Source in the same way through the same place at no
further charge. You need not require recipients to copy the
Corresponding Source along with the object code. If the place to
copy the object code is a network server, the Corresponding Source
may be on a different server (operated by you or a third party)
that supports equivalent copying facilities, provided you maintain
clear directions next to the object code saying where to find the
Corresponding Source. Regardless of what server hosts the
Corresponding Source, you remain obligated to ensure that it is
available for as long as needed to satisfy these requirements.
e) Convey the object code using peer-to-peer transmission, provided
you inform other peers where the object code and Corresponding
Source of the work are being offered to the general public at no
charge under subsection 6d.
A separable portion of the object code, whose source code is excluded
from the Corresponding Source as a System Library, need not be
included in conveying the object code work.
A "User Product" is either (1) a "consumer product", which means any
tangible personal property which is normally used for personal, family,
or household purposes, or (2) anything designed or sold for incorporation
into a dwelling. In determining whether a product is a consumer product,
doubtful cases shall be resolved in favor of coverage. For a particular
product received by a particular user, "normally used" refers to a
typical or common use of that class of product, regardless of the status
of the particular user or of the way in which the particular user
actually uses, or expects or is expected to use, the product. A product
is a consumer product regardless of whether the product has substantial
commercial, industrial or non-consumer uses, unless such uses represent
the only significant mode of use of the product.
"Installation Information" for a User Product means any methods,
procedures, authorization keys, or other information required to install
and execute modified versions of a covered work in that User Product from
a modified version of its Corresponding Source. The information must
suffice to ensure that the continued functioning of the modified object
code is in no case prevented or interfered with solely because
modification has been made.
If you convey an object code work under this section in, or with, or
specifically for use in, a User Product, and the conveying occurs as
part of a transaction in which the right of possession and use of the
User Product is transferred to the recipient in perpetuity or for a
fixed term (regardless of how the transaction is characterized), the
Corresponding Source conveyed under this section must be accompanied
by the Installation Information. But this requirement does not apply
if neither you nor any third party retains the ability to install
modified object code on the User Product (for example, the work has
been installed in ROM).
The requirement to provide Installation Information does not include a
requirement to continue to provide support service, warranty, or updates
for a work that has been modified or installed by the recipient, or for
the User Product in which it has been modified or installed. Access to a
network may be denied when the modification itself materially and
adversely affects the operation of the network or violates the rules and
protocols for communication across the network.
Corresponding Source conveyed, and Installation Information provided,
in accord with this section must be in a format that is publicly
documented (and with an implementation available to the public in
source code form), and must require no special password or key for
unpacking, reading or copying.
7. Additional Terms.
"Additional permissions" are terms that supplement the terms of this
License by making exceptions from one or more of its conditions.
Additional permissions that are applicable to the entire Program shall
be treated as though they were included in this License, to the extent
that they are valid under applicable law. If additional permissions
apply only to part of the Program, that part may be used separately
under those permissions, but the entire Program remains governed by
this License without regard to the additional permissions.
When you convey a copy of a covered work, you may at your option
remove any additional permissions from that copy, or from any part of
it. (Additional permissions may be written to require their own
removal in certain cases when you modify the work.) You may place
additional permissions on material, added by you to a covered work,
for which you have or can give appropriate copyright permission.
Notwithstanding any other provision of this License, for material you
add to a covered work, you may (if authorized by the copyright holders of
that material) supplement the terms of this License with terms:
a) Disclaiming warranty or limiting liability differently from the
terms of sections 15 and 16 of this License; or
b) Requiring preservation of specified reasonable legal notices or
author attributions in that material or in the Appropriate Legal
Notices displayed by works containing it; or
c) Prohibiting misrepresentation of the origin of that material, or
requiring that modified versions of such material be marked in
reasonable ways as different from the original version; or
d) Limiting the use for publicity purposes of names of licensors or
authors of the material; or
e) Declining to grant rights under trademark law for use of some
trade names, trademarks, or service marks; or
f) Requiring indemnification of licensors and authors of that
material by anyone who conveys the material (or modified versions of
it) with contractual assumptions of liability to the recipient, for
any liability that these contractual assumptions directly impose on
those licensors and authors.
All other non-permissive additional terms are considered "further
restrictions" within the meaning of section 10. If the Program as you
received it, or any part of it, contains a notice stating that it is
governed by this License along with a term that is a further
restriction, you may remove that term. If a license document contains
a further restriction but permits relicensing or conveying under this
License, you may add to a covered work material governed by the terms
of that license document, provided that the further restriction does
not survive such relicensing or conveying.
If you add terms to a covered work in accord with this section, you
must place, in the relevant source files, a statement of the
additional terms that apply to those files, or a notice indicating
where to find the applicable terms.
Additional terms, permissive or non-permissive, may be stated in the
form of a separately written license, or stated as exceptions;
the above requirements apply either way.
8. Termination.
You may not propagate or modify a covered work except as expressly
provided under this License. Any attempt otherwise to propagate or
modify it is void, and will automatically terminate your rights under
this License (including any patent licenses granted under the third
paragraph of section 11).
However, if you cease all violation of this License, then your
license from a particular copyright holder is reinstated (a)
provisionally, unless and until the copyright holder explicitly and
finally terminates your license, and (b) permanently, if the copyright
holder fails to notify you of the violation by some reasonable means
prior to 60 days after the cessation.
Moreover, your license from a particular copyright holder is
reinstated permanently if the copyright holder notifies you of the
violation by some reasonable means, this is the first time you have
received notice of violation of this License (for any work) from that
copyright holder, and you cure the violation prior to 30 days after
your receipt of the notice.
Termination of your rights under this section does not terminate the
licenses of parties who have received copies or rights from you under
this License. If your rights have been terminated and not permanently
reinstated, you do not qualify to receive new licenses for the same
material under section 10.
9. Acceptance Not Required for Having Copies.
You are not required to accept this License in order to receive or
run a copy of the Program. Ancillary propagation of a covered work
occurring solely as a consequence of using peer-to-peer transmission
to receive a copy likewise does not require acceptance. However,
nothing other than this License grants you permission to propagate or
modify any covered work. These actions infringe copyright if you do
not accept this License. Therefore, by modifying or propagating a
covered work, you indicate your acceptance of this License to do so.
10. Automatic Licensing of Downstream Recipients.
Each time you convey a covered work, the recipient automatically
receives a license from the original licensors, to run, modify and
propagate that work, subject to this License. You are not responsible
for enforcing compliance by third parties with this License.
An "entity transaction" is a transaction transferring control of an
organization, or substantially all assets of one, or subdividing an
organization, or merging organizations. If propagation of a covered
work results from an entity transaction, each party to that
transaction who receives a copy of the work also receives whatever
licenses to the work the party's predecessor in interest had or could
give under the previous paragraph, plus a right to possession of the
Corresponding Source of the work from the predecessor in interest, if
the predecessor has it or can get it with reasonable efforts.
You may not impose any further restrictions on the exercise of the
rights granted or affirmed under this License. For example, you may
not impose a license fee, royalty, or other charge for exercise of
rights granted under this License, and you may not initiate litigation
(including a cross-claim or counterclaim in a lawsuit) alleging that
any patent claim is infringed by making, using, selling, offering for
sale, or importing the Program or any portion of it.
11. Patents.
A "contributor" is a copyright holder who authorizes use under this
License of the Program or a work on which the Program is based. The
work thus licensed is called the contributor's "contributor version".
A contributor's "essential patent claims" are all patent claims
owned or controlled by the contributor, whether already acquired or
hereafter acquired, that would be infringed by some manner, permitted
by this License, of making, using, or selling its contributor version,
but do not include claims that would be infringed only as a
consequence of further modification of the contributor version. For
purposes of this definition, "control" includes the right to grant
patent sublicenses in a manner consistent with the requirements of
this License.
Each contributor grants you a non-exclusive, worldwide, royalty-free
patent license under the contributor's essential patent claims, to
make, use, sell, offer for sale, import and otherwise run, modify and
propagate the contents of its contributor version.
In the following three paragraphs, a "patent license" is any express
agreement or commitment, however denominated, not to enforce a patent
(such as an express permission to practice a patent or covenant not to
sue for patent infringement). To "grant" such a patent license to a
party means to make such an agreement or commitment not to enforce a
patent against the party.
If you convey a covered work, knowingly relying on a patent license,
and the Corresponding Source of the work is not available for anyone
to copy, free of charge and under the terms of this License, through a
publicly available network server or other readily accessible means,
then you must either (1) cause the Corresponding Source to be so
available, or (2) arrange to deprive yourself of the benefit of the
patent license for this particular work, or (3) arrange, in a manner
consistent with the requirements of this License, to extend the patent
license to downstream recipients. "Knowingly relying" means you have
actual knowledge that, but for the patent license, your conveying the
covered work in a country, or your recipient's use of the covered work
in a country, would infringe one or more identifiable patents in that
country that you have reason to believe are valid.
If, pursuant to or in connection with a single transaction or
arrangement, you convey, or propagate by procuring conveyance of, a
covered work, and grant a patent license to some of the parties
receiving the covered work authorizing them to use, propagate, modify
or convey a specific copy of the covered work, then the patent license
you grant is automatically extended to all recipients of the covered
work and works based on it.
A patent license is "discriminatory" if it does not include within
the scope of its coverage, prohibits the exercise of, or is
conditioned on the non-exercise of one or more of the rights that are
specifically granted under this License. You may not convey a covered
work if you are a party to an arrangement with a third party that is
in the business of distributing software, under which you make payment
to the third party based on the extent of your activity of conveying
the work, and under which the third party grants, to any of the
parties who would receive the covered work from you, a discriminatory
patent license (a) in connection with copies of the covered work
conveyed by you (or copies made from those copies), or (b) primarily
for and in connection with specific products or compilations that
contain the covered work, unless you entered into that arrangement,
or that patent license was granted, prior to 28 March 2007.
Nothing in this License shall be construed as excluding or limiting
any implied license or other defenses to infringement that may
otherwise be available to you under applicable patent law.
12. No Surrender of Others' Freedom.
If conditions are imposed on you (whether by court order, agreement or
otherwise) that contradict the conditions of this License, they do not
excuse you from the conditions of this License. If you cannot convey a
covered work so as to satisfy simultaneously your obligations under this
License and any other pertinent obligations, then as a consequence you may
not convey it at all. For example, if you agree to terms that obligate you
to collect a royalty for further conveying from those to whom you convey
the Program, the only way you could satisfy both those terms and this
License would be to refrain entirely from conveying the Program.
13. Use with the GNU Affero General Public License.
Notwithstanding any other provision of this License, you have
permission to link or combine any covered work with a work licensed
under version 3 of the GNU Affero General Public License into a single
combined work, and to convey the resulting work. The terms of this
License will continue to apply to the part which is the covered work,
but the special requirements of the GNU Affero General Public License,
section 13, concerning interaction through a network will apply to the
combination as such.
14. Revised Versions of this License.
The Free Software Foundation may publish revised and/or new versions of
the GNU General Public License from time to time. Such new versions will
be similar in spirit to the present version, but may differ in detail to
address new problems or concerns.
Each version is given a distinguishing version number. If the
Program specifies that a certain numbered version of the GNU General
Public License "or any later version" applies to it, you have the
option of following the terms and conditions either of that numbered
version or of any later version published by the Free Software
Foundation. If the Program does not specify a version number of the
GNU General Public License, you may choose any version ever published
by the Free Software Foundation.
If the Program specifies that a proxy can decide which future
versions of the GNU General Public License can be used, that proxy's
public statement of acceptance of a version permanently authorizes you
to choose that version for the Program.
Later license versions may give you additional or different
permissions. However, no additional obligations are imposed on any
author or copyright holder as a result of your choosing to follow a
later version.
15. Disclaimer of Warranty.
THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
16. Limitation of Liability.
IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
SUCH DAMAGES.
17. Interpretation of Sections 15 and 16.
If the disclaimer of warranty and limitation of liability provided
above cannot be given local legal effect according to their terms,
reviewing courts shall apply local law that most closely approximates
an absolute waiver of all civil liability in connection with the
Program, unless a warranty or assumption of liability accompanies a
copy of the Program in return for a fee.
END OF TERMS AND CONDITIONS
How to Apply These Terms to Your New Programs
If you develop a new program, and you want it to be of the greatest
possible use to the public, the best way to achieve this is to make it
free software which everyone can redistribute and change under these terms.
To do so, attach the following notices to the program. It is safest
to attach them to the start of each source file to most effectively
state the exclusion of warranty; and each file should have at least
the "copyright" line and a pointer to where the full notice is found.
<one line to give the program's name and a brief idea of what it does.>
Copyright (C) <year> <name of author>
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see <http://www.gnu.org/licenses/>.
Also add information on how to contact you by electronic and paper mail.
If the program does terminal interaction, make it output a short
notice like this when it starts in an interactive mode:
<program> Copyright (C) <year> <name of author>
This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
This is free software, and you are welcome to redistribute it
under certain conditions; type `show c' for details.
The hypothetical commands `show w' and `show c' should show the appropriate
parts of the General Public License. Of course, your program's commands
might be different; for a GUI interface, you would use an "about box".
You should also get your employer (if you work as a programmer) or school,
if any, to sign a "copyright disclaimer" for the program, if necessary.
For more information on this, and how to apply and follow the GNU GPL, see
<http://www.gnu.org/licenses/>.
The GNU General Public License does not permit incorporating your program
into proprietary programs. If your program is a subroutine library, you
may consider it more useful to permit linking proprietary applications with
the library. If this is what you want to do, use the GNU Lesser General
Public License instead of this License. But first, please read
<http://www.gnu.org/philosophy/why-not-lgpl.html>.

View File

@ -1,3 +0,0 @@
from .api import StandaloneSATRNRunner
from .version import __version__
from .factory import __hub_available_versions__

View File

@ -1,106 +0,0 @@
import torch
import torch.nn as nn
from colorama import Fore, Style
from .converter import AttnConvertor
from .backbone import ResNetABI
from .encoder import SatrnEncoder
from .decoder import NRTRDecoder
from .transform import DataPipelineSATRN
from .fp16_utils import patch_norm_fp32
from .factory import _get as get_version
class SATRN(nn.Module):
"""Standalone implementation for SATRN encode-decode recognizer."""
def __init__(self,
version,
return_confident=False,
device='cpu',
max_seq_len_overwrite=None):
super().__init__()
checkpoint = get_version(version)
pt = torch.load(checkpoint, 'cpu')
if device == 'cpu':
print(Fore.RED + 'Warning: You are using CPU inference method. Init with device=cuda:<id> to run with CUDA method.' + Style.RESET_ALL)
self.pipeline = DataPipelineSATRN(**pt['pipeline_args'], device=device)
# Convertor
self.label_convertor = AttnConvertor(**pt['label_convertor_args'], return_confident=return_confident)
# Backbone
self.backbone = ResNetABI(**pt['backbone_args'])
# Encoder module
self.encoder = SatrnEncoder(**pt['encoder_args'])
# Decoder module
decoder_max_seq_len = max_seq_len_overwrite if max_seq_len_overwrite is not None else pt['max_seq_len']
self.decoder = NRTRDecoder(
**pt['decoder_args'],
max_seq_len=decoder_max_seq_len,
num_classes=self.label_convertor.num_classes(),
start_idx=self.label_convertor.start_idx,
padding_idx=self.label_convertor.padding_idx,
return_confident=return_confident,
end_idx=self.label_convertor.end_idx
)
self.load_state_dict(pt['state_dict'], strict=True)
print(f'Text recognition from version {version}.')
if device != 'cpu':
self = self.to(device)
self = self.half()
patch_norm_fp32(self)
self.eval()
for param in self.parameters():
param.requires_grad = False
def extract_feat(self, img):
x = self.backbone(img)
return x
def forward(self, img):
"""Test function with test time augmentation.
Args:
imgs (torch.Tensor): Image input tensor.
Returns:
list[str]: Text label result of each image.
"""
img = self.pipeline(img)
feat = self.extract_feat(img)
out_enc = self.encoder(feat)
out_dec = self.decoder(out_enc).cpu().numpy()
label_strings = self.label_convertor(out_dec)
return label_strings
class StandaloneSATRNRunner:
def __init__(self,
version,
return_confident,
device='cpu',
max_seq_len_overwrite=None):
self.device = device
self.model = SATRN(version=version,
return_confident=return_confident,
device=self.device,
max_seq_len_overwrite=max_seq_len_overwrite)
def __call__(self, imgs):
results = self.model(imgs)
return results

View File

@ -1,159 +0,0 @@
import torch.nn as nn
def conv3x3(in_planes, out_planes, stride=1):
"""3x3 convolution with padding."""
return nn.Conv2d(
in_planes,
out_planes,
kernel_size=3,
stride=stride,
padding=1,
bias=False)
def conv1x1(in_planes, out_planes):
"""1x1 convolution with padding."""
return nn.Conv2d(
in_planes, out_planes, kernel_size=1, stride=1, padding=0, bias=False)
class BasicBlock(nn.Module):
expansion = 1
def __init__(self,
inplanes,
planes,
stride=1,
downsample=None,
use_conv1x1=False):
super(BasicBlock, self).__init__()
if use_conv1x1:
self.conv1 = conv1x1(inplanes, planes)
self.conv2 = conv3x3(planes, planes * self.expansion, stride)
else:
self.conv1 = conv3x3(inplanes, planes, stride)
self.conv2 = conv3x3(planes, planes * self.expansion)
self.planes = planes
self.bn1 = nn.BatchNorm2d(planes)
self.relu = nn.ReLU(inplace=True)
self.bn2 = nn.BatchNorm2d(planes * self.expansion)
self.downsample = downsample
self.stride = stride
def forward(self, x):
residual = x
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
out = self.conv2(out)
out = self.bn2(out)
if self.downsample is not None:
residual = self.downsample(x)
out += residual
out = self.relu(out)
return out
class ResNetABI(nn.Module):
"""Implement ResNet backbone for text recognition, modified from `ResNet.
<https://arxiv.org/pdf/1512.03385.pdf>`_ and
`<https://github.com/FangShancheng/ABINet>`_
Args:
in_channels (int): Number of channels of input image tensor.
stem_channels (int): Number of stem channels.
base_channels (int): Number of base channels.
arch_settings (list[int]): List of BasicBlock number for each stage.
strides (Sequence[int]): Strides of the first block of each stage.
out_indices (None | Sequence[int]): Indices of output stages. If not
specified, only the last stage will be returned.
last_stage_pool (bool): If True, add `MaxPool2d` layer to last stage.
"""
def __init__(self,
in_channels=3,
stem_channels=32,
base_channels=32,
arch_settings=[3, 4, 6, 6, 3],
strides=[2, 1, 2, 1, 1],
out_indices=None,
last_stage_pool=False):
super().__init__()
self.out_indices = out_indices
self.last_stage_pool = last_stage_pool
self.block = BasicBlock
self.inplanes = stem_channels
self._make_stem_layer(in_channels, stem_channels)
self.res_layers = []
planes = base_channels
for i, num_blocks in enumerate(arch_settings):
stride = strides[i]
res_layer = self._make_layer(
block=self.block,
inplanes=self.inplanes,
planes=planes,
blocks=num_blocks,
stride=stride)
self.inplanes = planes * self.block.expansion
planes *= 2
layer_name = f'layer{i + 1}'
self.add_module(layer_name, res_layer)
self.res_layers.append(layer_name)
def _make_layer(self, block, inplanes, planes, blocks, stride=1):
layers = []
downsample = None
if stride != 1 or inplanes != planes:
downsample = nn.Sequential(
nn.Conv2d(inplanes, planes, 1, stride, bias=False),
nn.BatchNorm2d(planes),
)
layers.append(
block(
inplanes,
planes,
use_conv1x1=True,
stride=stride,
downsample=downsample))
inplanes = planes
for _ in range(1, blocks):
layers.append(block(inplanes, planes, use_conv1x1=True))
return nn.Sequential(*layers)
def _make_stem_layer(self, in_channels, stem_channels):
self.conv1 = nn.Conv2d(
in_channels, stem_channels, kernel_size=3, stride=1, padding=1)
self.bn1 = nn.BatchNorm2d(stem_channels)
self.relu1 = nn.ReLU(inplace=True)
def forward(self, x):
"""
Args:
x (Tensor): Image tensor of shape :math:`(N, 3, H, W)`.
Returns:
Tensor or list[Tensor]: Feature tensor. Its shape depends on
ResNetABI's config. It can be a list of feature outputs at specific
layers if ``out_indices`` is specified.
"""
x = self.conv1(x)
x = self.bn1(x)
x = self.relu1(x)
outs = []
for i, layer_name in enumerate(self.res_layers):
res_layer = getattr(self, layer_name)
x = res_layer(x)
if self.out_indices and i in self.out_indices:
outs.append(x)
return tuple(outs) if self.out_indices else x

View File

@ -1,173 +0,0 @@
import warnings
import torch.nn as nn
from torch.nn.modules.instancenorm import _InstanceNorm
from torch.nn.modules.batchnorm import _BatchNorm
from mmcv.cnn import build_padding_layer, build_conv_layer, build_norm_layer, build_activation_layer
class ConvModule(nn.Module):
"""A conv block that bundles conv/norm/activation layers.
This block simplifies the usage of convolution layers, which are commonly
used with a norm layer (e.g., BatchNorm) and activation layer (e.g., ReLU).
It is based upon three build methods: `build_conv_layer()`,
`build_norm_layer()` and `build_activation_layer()`.
Besides, we add some additional features in this module.
1. Automatically set `bias` of the conv layer.
2. Spectral norm is supported.
3. More padding modes are supported. Before PyTorch 1.5, nn.Conv2d only
supports zero and circular padding, and we add "reflect" padding mode.
Args:
in_channels (int): Number of channels in the input feature map.
Same as that in ``nn._ConvNd``.
out_channels (int): Number of channels produced by the convolution.
Same as that in ``nn._ConvNd``.
kernel_size (int | tuple[int]): Size of the convolving kernel.
Same as that in ``nn._ConvNd``.
stride (int | tuple[int]): Stride of the convolution.
Same as that in ``nn._ConvNd``.
padding (int | tuple[int]): Zero-padding added to both sides of
the input. Same as that in ``nn._ConvNd``.
dilation (int | tuple[int]): Spacing between kernel elements.
Same as that in ``nn._ConvNd``.
groups (int): Number of blocked connections from input channels to
output channels. Same as that in ``nn._ConvNd``.
bias (bool | str): If specified as `auto`, it will be decided by the
norm_cfg. Bias will be set as True if `norm_cfg` is None, otherwise
False. Default: "auto".
conv_cfg (dict): Config dict for convolution layer. Default: None,
which means using conv2d.
norm_cfg (dict): Config dict for normalization layer. Default: None.
act_cfg (dict): Config dict for activation layer.
Default: dict(type='ReLU').
inplace (bool): Whether to use inplace mode for activation.
Default: True.
with_spectral_norm (bool): Whether use spectral norm in conv module.
Default: False.
padding_mode (str): If the `padding_mode` has not been supported by
current `Conv2d` in PyTorch, we will use our own padding layer
instead. Currently, we support ['zeros', 'circular'] with official
implementation and ['reflect'] with our own implementation.
Default: 'zeros'.
order (tuple[str]): The order of conv/norm/activation layers. It is a
sequence of "conv", "norm" and "act". Common examples are
("conv", "norm", "act") and ("act", "conv", "norm").
Default: ('conv', 'norm', 'act').
"""
_abbr_ = 'conv_block'
def __init__(self,
in_channels,
out_channels,
kernel_size,
stride=1,
padding=0,
dilation=1,
groups=1,
bias='auto',
conv_cfg=None,
norm_cfg=None,
act_cfg=dict(type='ReLU'),
inplace=True,
with_spectral_norm=False,
padding_mode='zeros',
order=('conv', 'norm', 'act')):
super(ConvModule, self).__init__()
assert conv_cfg is None or isinstance(conv_cfg, dict)
assert norm_cfg is None or isinstance(norm_cfg, dict)
assert act_cfg is None or isinstance(act_cfg, dict)
official_padding_mode = ['zeros', 'circular']
self.conv_cfg = conv_cfg
self.norm_cfg = norm_cfg
self.act_cfg = act_cfg
self.inplace = inplace
self.with_spectral_norm = with_spectral_norm
self.with_explicit_padding = padding_mode not in official_padding_mode
self.order = order
assert isinstance(self.order, tuple) and len(self.order) == 3
assert set(order) == set(['conv', 'norm', 'act'])
self.with_norm = norm_cfg is not None
self.with_activation = act_cfg is not None
# if the conv layer is before a norm layer, bias is unnecessary.
if bias == 'auto':
bias = not self.with_norm
self.with_bias = bias
if self.with_explicit_padding:
pad_cfg = dict(type=padding_mode)
self.padding_layer = build_padding_layer(pad_cfg, padding)
# reset padding to 0 for conv module
conv_padding = 0 if self.with_explicit_padding else padding
# build convolution layer
self.conv = build_conv_layer(
conv_cfg,
in_channels,
out_channels,
kernel_size,
stride=stride,
padding=conv_padding,
dilation=dilation,
groups=groups,
bias=bias)
# export the attributes of self.conv to a higher level for convenience
self.in_channels = self.conv.in_channels
self.out_channels = self.conv.out_channels
self.kernel_size = self.conv.kernel_size
self.stride = self.conv.stride
self.padding = padding
self.dilation = self.conv.dilation
self.transposed = self.conv.transposed
self.output_padding = self.conv.output_padding
self.groups = self.conv.groups
if self.with_spectral_norm:
self.conv = nn.utils.spectral_norm(self.conv)
# build normalization layers
if self.with_norm:
# norm layer is after conv layer
if order.index('norm') > order.index('conv'):
norm_channels = out_channels
else:
norm_channels = in_channels
self.norm_name, norm = build_norm_layer(norm_cfg, norm_channels)
self.add_module(self.norm_name, norm)
if self.with_bias:
if isinstance(norm, (_BatchNorm, _InstanceNorm)):
warnings.warn(
'Unnecessary conv bias before batch/instance norm')
else:
self.norm_name = None
# build activation layer
if self.with_activation:
act_cfg_ = act_cfg.copy()
# nn.Tanh has no 'inplace' argument
if act_cfg_['type'] not in [
'Tanh', 'PReLU', 'Sigmoid', 'HSigmoid', 'Swish'
]:
act_cfg_.setdefault('inplace', inplace)
self.activate = build_activation_layer(act_cfg_)
@property
def norm(self):
if self.norm_name:
return getattr(self, self.norm_name)
else:
return None
def forward(self, x, activate=True, norm=True):
for layer in self.order:
if layer == 'conv':
if self.with_explicit_padding:
x = self.padding_layer(x)
x = self.conv(x)
elif layer == 'norm' and norm and self.with_norm:
x = self.norm(x)
elif layer == 'act' and activate and self.with_activation:
x = self.activate(x)
return x

View File

@ -1,152 +0,0 @@
import torch
import numpy as np
class BaseConvertor:
"""Convert between text, index and tensor for text recognize pipeline.
Args:
dict_type (str): Type of dict, options are 'DICT36', 'DICT37', 'DICT90'
and 'DICT91'.
dict_file (None|str): Character dict file path. If not none,
the dict_file is of higher priority than dict_type.
dict_list (None|list[str]): Character list. If not none, the list
is of higher priority than dict_type, but lower than dict_file.
"""
start_idx = end_idx = padding_idx = 0
unknown_idx = None
lower = False
DICT36 = tuple('0123456789abcdefghijklmnopqrstuvwxyz')
DICT63 = tuple('0123456789abcdefghijklmnopqrstuvwxyz'
'ABCDEFGHIJKLMNOPQRSTUVWXYZ')
DICT90 = tuple('0123456789abcdefghijklmnopqrstuvwxyz'
'ABCDEFGHIJKLMNOPQRSTUVWXYZ!"#$%&\'()'
'*+,-./:;<=>?@[\\]_`~')
DICT131 = tuple('0123456789abcdefghijklmnopqrstuvwxyz'
'!"#$%&\'()'
'*+,-./:;<=>?@[\\]_`~'
'ạảãàáâậầấẩẫăắằặẳẵóòọõỏôộổỗồốơờớợởỡéèẻẹẽêếềệểễúùụủũưựữửừứíìịỉĩýỳỷỵỹđ')
DICT224 = tuple('0123456789abcdefghijklmnopqrstuvwxyz'
'ABCDEFGHIJKLMNOPQRSTUVWXYZ!"#$%&\'()'
'*+,-./:;<=>?@[\\]_`~{}|^ ̂'
'ạảãàáâậầấẩẫăắằặẳẵóòọõỏôộổỗồốơờớợởỡéèẻẹẽêếềệểễúùụủũưựữửừứíìịỉĩýỳỷỵỹđ'
'ẠẢÃÀÁÂẬẦẤẨẪĂẮẰẶẲẴÓÒỌÕỎÔỘỔỖỒỐƠỜỚỢỞỠÉÈẺẸẼÊẾỀỆỂỄÚÙỤỦŨƯỰỮỬỪỨÍÌỊỈĨÝỲỶỴỸĐ✪')
def __init__(self, dict_type='DICT90'):
assert dict_type in ('DICT36', 'DICT63', 'DICT90', 'DICT131', 'DICT224')
self.idx2char = []
if dict_type == 'DICT36':
self.idx2char = list(self.DICT36)
elif dict_type == 'DICT63':
self.idx2char = list(self.DICT63)
elif dict_type == 'DICT90':
self.idx2char = list(self.DICT90)
elif dict_type == 'DICT131':
self.idx2char = list(self.DICT131)
elif dict_type == 'DICT224':
self.idx2char = list(self.DICT224)
else:
raise ('Dictonary not implemented')
assert len(set(self.idx2char)) == len(self.idx2char), \
'Invalid dictionary: Has duplicated characters.'
self.char2idx = {char: idx for idx, char in enumerate(self.idx2char)}
def num_classes(self):
"""Number of output classes."""
return len(self.idx2char)
class AttnConvertor(BaseConvertor):
"""Convert between text, index and tensor for encoder-decoder based
pipeline.
Args:
dict_type (str): Type of dict, should be one of {'DICT36', 'DICT90'}.
dict_file (None|str): Character dict file path. If not none,
higher priority than dict_type.
dict_list (None|list[str]): Character list. If not none, higher
priority than dict_type, but lower than dict_file.
with_unknown (bool): If True, add `UKN` token to class.
max_seq_len (int): Maximum sequence length of label.
lower (bool): If True, convert original string to lower case.
start_end_same (bool): Whether use the same index for
start and end token or not. Default: True.
"""
def __init__(self,
dict_type='DICT90',
with_unknown=True,
max_seq_len=40,
lower=False,
start_end_same=True,
return_confident=False,
**kwargs):
super().__init__(dict_type)
assert isinstance(with_unknown, bool)
assert isinstance(max_seq_len, int)
assert isinstance(lower, bool)
self.with_unknown = with_unknown
self.max_seq_len = max_seq_len
self.lower = lower
self.start_end_same = start_end_same
self.return_confident = return_confident
self.update_dict()
def update_dict(self):
start_end_token = '<BOS/EOS>'
unknown_token = '<UKN>'
padding_token = '<PAD>'
# unknown
self.unknown_idx = None
if self.with_unknown:
self.idx2char.append(unknown_token)
self.unknown_idx = len(self.idx2char) - 1
# BOS/EOS
self.idx2char.append(start_end_token)
self.start_idx = len(self.idx2char) - 1
if not self.start_end_same:
self.idx2char.append(start_end_token)
self.end_idx = len(self.idx2char) - 1
# padding
self.idx2char.append(padding_token)
self.padding_idx = len(self.idx2char) - 1
# update char2idx
self.char2idx = {}
for idx, char in enumerate(self.idx2char):
self.char2idx[char] = idx
def __call__(self, indexes):
strings = []
confidents = []
if self.return_confident:
b,sq,_ = indexes.shape
for idx in range(b):
index = indexes[idx, :, :]
chars = index.argmax(-1)
confident = index.max(-1)
i = 0
while i < sq and chars[i] != self.end_idx and chars[i] != self.padding_idx: i += 1
chars = chars[:i]
confident = confident[:i].min()
string = [self.idx2char[i] for i in chars]
strings.append(''.join(string))
confidents.append(confident)
return strings, confidents
else:
for index in indexes:
i, l = 0, len(index)
while i < l and index[i] != self.end_idx and index[i] != self.padding_idx: i += 1
index = index[:i]
string = [self.idx2char[i] for i in index]
strings.append(''.join(string))
return strings

View File

@ -1,278 +0,0 @@
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
import numpy as np
from .encoder import MultiHeadAttention
class PositionwiseFeedForward(nn.Module):
"""Two-layer feed-forward module.
Args:
d_in (int): The dimension of the input for feedforward
network model.
d_hid (int): The dimension of the feedforward
network model.
dropout (float): Dropout layer on feedforward output.
act_cfg (dict): Activation cfg for feedforward module.
"""
def __init__(self, d_in, d_hid, dropout=0.1):
super().__init__()
self.w_1 = nn.Linear(d_in, d_hid)
self.w_2 = nn.Linear(d_hid, d_in)
self.act = nn.GELU()
self.dropout = nn.Dropout(dropout)
def forward(self, x):
x = self.w_1(x)
x = self.act(x)
x = self.w_2(x)
x = self.dropout(x)
return x
class PositionalEncoding(nn.Module):
"""Fixed positional encoding with sine and cosine functions."""
def __init__(self, d_hid=512, n_position=200):
super().__init__()
# Not a parameter
# Position table of shape (1, n_position, d_hid)
self.register_buffer(
'position_table',
self._get_sinusoid_encoding_table(n_position, d_hid))
def _get_sinusoid_encoding_table(self, n_position, d_hid):
"""Sinusoid position encoding table."""
denominator = torch.Tensor([
1.0 / np.power(10000, 2 * (hid_j // 2) / d_hid)
for hid_j in range(d_hid)
])
denominator = denominator.view(1, -1)
pos_tensor = torch.arange(n_position).unsqueeze(-1).float()
sinusoid_table = pos_tensor * denominator
sinusoid_table[:, 0::2] = torch.sin(sinusoid_table[:, 0::2])
sinusoid_table[:, 1::2] = torch.cos(sinusoid_table[:, 1::2])
return sinusoid_table.unsqueeze(0)
def forward(self, x):
"""
Args:
x (Tensor): Tensor of shape (batch_size, pos_len, d_hid, ...)
"""
self.device = x.device
x = x + self.position_table[:, :x.size(1)].clone().detach()
return x
class TFDecoderLayer(nn.Module):
"""Transformer Decoder Layer.
Args:
d_model (int): The number of expected features
in the decoder inputs (default=512).
d_inner (int): The dimension of the feedforward
network model (default=256).
n_head (int): The number of heads in the
multiheadattention models (default=8).
d_k (int): Total number of features in key.
d_v (int): Total number of features in value.
dropout (float): Dropout layer on attn_output_weights.
qkv_bias (bool): Add bias in projection layer. Default: False.
act_cfg (dict): Activation cfg for feedforward module.
operation_order (tuple[str]): The execution order of operation
in transformer. Such as ('self_attn', 'norm', 'enc_dec_attn',
'norm', 'ffn', 'norm') or ('norm', 'self_attn', 'norm',
'enc_dec_attn', 'norm', 'ffn').
Default:None.
"""
def __init__(self,
d_model=512,
d_inner=256,
n_head=8,
d_k=64,
d_v=64,
dropout=0.1,
qkv_bias=False):
super().__init__()
self.norm1 = nn.LayerNorm(d_model)
self.norm2 = nn.LayerNorm(d_model)
self.norm3 = nn.LayerNorm(d_model)
self.self_attn = MultiHeadAttention(
n_head, d_model, d_k, d_v, dropout=dropout, qkv_bias=qkv_bias)
self.enc_attn = MultiHeadAttention(
n_head, d_model, d_k, d_v, dropout=dropout, qkv_bias=qkv_bias)
self.mlp = PositionwiseFeedForward(
d_model, d_inner, dropout=dropout)
def forward(self,
dec_input,
enc_output,
self_attn_mask=None,
dec_enc_attn_mask=None):
dec_input_norm = self.norm1(dec_input)
dec_attn_out = self.self_attn(dec_input_norm, dec_input_norm,
dec_input_norm, self_attn_mask)
dec_attn_out += dec_input
enc_dec_attn_in = self.norm2(dec_attn_out)
enc_dec_attn_out = self.enc_attn(enc_dec_attn_in, enc_output,
enc_output, dec_enc_attn_mask)
enc_dec_attn_out += dec_attn_out
mlp_out = self.mlp(self.norm3(enc_dec_attn_out))
mlp_out += enc_dec_attn_out
return mlp_out
class NRTRDecoder(nn.Module):
"""Transformer Decoder block with self attention mechanism.
Args:
n_layers (int): Number of attention layers.
d_embedding (int): Language embedding dimension.
n_head (int): Number of parallel attention heads.
d_k (int): Dimension of the key vector.
d_v (int): Dimension of the value vector.
d_model (int): Dimension :math:`D_m` of the input from previous model.
d_inner (int): Hidden dimension of feedforward layers.
n_position (int): Length of the positional encoding vector. Must be
greater than ``max_seq_len``.
dropout (float): Dropout rate.
num_classes (int): Number of output classes :math:`C`.
max_seq_len (int): Maximum output sequence length :math:`T`.
start_idx (int): The index of `<SOS>`.
padding_idx (int): The index of `<PAD>`.
init_cfg (dict or list[dict], optional): Initialization configs.
Warning:
This decoder will not predict the final class which is assumed to be
`<PAD>`. Therefore, its output size is always :math:`C - 1`. `<PAD>`
is also ignored by loss as specified in
:obj:`mmocr.models.textrecog.recognizer.EncodeDecodeRecognizer`.
"""
def __init__(self,
n_layers=6,
d_embedding=512,
n_head=8,
d_k=64,
d_v=64,
d_model=512,
d_inner=256,
n_position=200,
dropout=0.1,
num_classes=93,
max_seq_len=40,
start_idx=1,
padding_idx=92,
return_confident=False,
end_idx=None,
**kwargs):
super().__init__()
self.padding_idx = padding_idx
self.start_idx = start_idx
self.max_seq_len = max_seq_len
self.trg_word_emb = nn.Embedding(
num_classes, d_embedding, padding_idx=padding_idx)
self.position_enc = PositionalEncoding(
d_embedding, n_position=n_position)
self.layer_stack = nn.ModuleList([
TFDecoderLayer(
d_model, d_inner, n_head, d_k, d_v, dropout=dropout, **kwargs)
for _ in range(n_layers)
])
self.layer_norm = nn.LayerNorm(d_model, eps=1e-6)
pred_num_class = num_classes - 1 # ignore padding_idx
self.classifier = nn.Linear(d_model, pred_num_class)
self.return_confident = return_confident
self.end_idx = end_idx
@staticmethod
def get_pad_mask(seq, pad_idx):
return (seq != pad_idx).unsqueeze(-2)
@staticmethod
def get_subsequent_mask(seq):
"""For masking out the subsequent info."""
len_s = seq.size(1)
subsequent_mask = 1 - torch.triu(
torch.ones((len_s, len_s), device=seq.device), diagonal=1)
subsequent_mask = subsequent_mask.unsqueeze(0).bool()
return subsequent_mask
def _attention(self, trg_seq, src, src_mask=None):
trg_embedding = self.trg_word_emb(trg_seq)
trg_pos_encoded = self.position_enc(trg_embedding)
trg_mask = self.get_pad_mask(
trg_seq,
pad_idx=self.padding_idx) & self.get_subsequent_mask(trg_seq)
output = trg_pos_encoded
for dec_layer in self.layer_stack:
output = dec_layer(
output,
src,
self_attn_mask=trg_mask,
dec_enc_attn_mask=src_mask)
output = self.layer_norm(output)
return output
def _get_mask(self, logit):
N, T, _ = logit.size()
mask = logit.new_ones((N, T))
return mask
def forward(self, out_enc):
src_mask = self._get_mask(out_enc)
N = out_enc.size(0)
init_target_seq = torch.full((N, self.max_seq_len + 1),
self.padding_idx,
device=out_enc.device,
dtype=torch.long)
# bsz * seq_len
init_target_seq[:, 0] = self.start_idx
outputs = []
for step in range(0, self.max_seq_len):
decoder_output = self._attention(
trg_seq=init_target_seq,
src=out_enc,
src_mask=src_mask)
if self.return_confident:
step_result = torch.softmax(self.classifier(decoder_output[:, step, :]), -1)
next_step_init = step_result.argmax(-1)
init_target_seq[:, step + 1] = next_step_init
if next_step_init.min() >= self.end_idx:
break
else:
step_result = self.classifier(decoder_output[:, step, :]).argmax(-1)
init_target_seq[:, step + 1] = step_result
if step_result.min() >= self.end_idx:
break
outputs.append(step_result)
outputs = torch.stack(outputs, dim=1)
return outputs

View File

@ -1,317 +0,0 @@
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import math
from .conv import ConvModule
class LocalityAwareFeedforward(nn.Module):
"""Locality-aware feedforward layer in SATRN, see `SATRN.
<https://arxiv.org/abs/1910.04396>`_
"""
def __init__(self,
d_in,
d_hid,
dropout=0.1):
super().__init__()
self.conv1 = ConvModule(
d_in,
d_hid,
kernel_size=1,
padding=0,
bias=False,
norm_cfg=dict(type='BN'),
act_cfg=dict(type='ReLU'))
self.depthwise_conv = ConvModule(
d_hid,
d_hid,
kernel_size=3,
padding=1,
bias=False,
groups=d_hid,
norm_cfg=dict(type='BN'),
act_cfg=dict(type='ReLU'))
self.conv2 = ConvModule(
d_hid,
d_in,
kernel_size=1,
padding=0,
bias=False,
norm_cfg=dict(type='BN'),
act_cfg=dict(type='ReLU'))
def forward(self, x):
x = self.conv1(x)
x = self.depthwise_conv(x)
x = self.conv2(x)
return x
class ScaledDotProductAttention(nn.Module):
"""Scaled Dot-Product Attention Module. This code is adopted from
https://github.com/jadore801120/attention-is-all-you-need-pytorch.
Args:
temperature (float): The scale factor for softmax input.
attn_dropout (float): Dropout layer on attn_output_weights.
"""
def __init__(self, temperature, attn_dropout=0.1):
super().__init__()
self.temperature = temperature
self.dropout = nn.Dropout(attn_dropout)
def forward(self, q, k, v, mask=None):
attn = torch.matmul(q / self.temperature, k.transpose(2, 3))
if mask is not None:
attn = attn.masked_fill(mask == 0, float('-inf'))
attn = self.dropout(F.softmax(attn, dim=-1))
output = torch.matmul(attn, v)
return output, attn
class MultiHeadAttention(nn.Module):
"""Multi-Head Attention module.
Args:
n_head (int): The number of heads in the
multiheadattention models (default=8).
d_model (int): The number of expected features
in the decoder inputs (default=512).
d_k (int): Total number of features in key.
d_v (int): Total number of features in value.
dropout (float): Dropout layer on attn_output_weights.
qkv_bias (bool): Add bias in projection layer. Default: False.
"""
def __init__(self,
n_head=8,
d_model=512,
d_k=64,
d_v=64,
dropout=0.1,
qkv_bias=False):
super().__init__()
self.n_head = n_head
self.d_k = d_k
self.d_v = d_v
self.dim_k = n_head * d_k
self.dim_v = n_head * d_v
self.linear_q = nn.Linear(self.dim_k, self.dim_k, bias=qkv_bias)
self.linear_k = nn.Linear(self.dim_k, self.dim_k, bias=qkv_bias)
self.linear_v = nn.Linear(self.dim_v, self.dim_v, bias=qkv_bias)
self.attention = ScaledDotProductAttention(d_k**0.5, dropout)
self.fc = nn.Linear(self.dim_v, d_model, bias=qkv_bias)
self.proj_drop = nn.Dropout(dropout)
def forward(self, q, k, v, mask=None):
batch_size, len_q, _ = q.size()
_, len_k, _ = k.size()
q = self.linear_q(q).view(batch_size, len_q, self.n_head, self.d_k)
k = self.linear_k(k).view(batch_size, len_k, self.n_head, self.d_k)
v = self.linear_v(v).view(batch_size, len_k, self.n_head, self.d_v)
q, k, v = q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2)
if mask is not None:
if mask.dim() == 3:
mask = mask.unsqueeze(1)
elif mask.dim() == 2:
mask = mask.unsqueeze(1).unsqueeze(1)
attn_out, _ = self.attention(q, k, v, mask=mask)
attn_out = attn_out.transpose(1, 2).contiguous().view(
batch_size, len_q, self.dim_v)
attn_out = self.fc(attn_out)
attn_out = self.proj_drop(attn_out)
return attn_out
class SatrnEncoderLayer(nn.Module):
""""""
def __init__(self,
d_model=512,
d_inner=512,
n_head=8,
d_k=64,
d_v=64,
dropout=0.1,
qkv_bias=False):
super().__init__()
self.norm1 = nn.LayerNorm(d_model)
self.attn = MultiHeadAttention(
n_head, d_model, d_k, d_v, qkv_bias=qkv_bias, dropout=dropout)
self.norm2 = nn.LayerNorm(d_model)
self.feed_forward = LocalityAwareFeedforward(
d_model, d_inner, dropout=dropout)
def forward(self, x, h, w, mask=None):
n, hw, c = x.size()
residual = x
x = self.norm1(x)
x = residual + self.attn(x, x, x, mask)
residual = x
x = self.norm2(x)
x = x.transpose(1, 2).contiguous().view(n, c, h, w)
x = self.feed_forward(x)
x = x.view(n, c, hw).transpose(1, 2)
x = residual + x
return x
class Adaptive2DPositionalEncoding(nn.Module):
"""Implement Adaptive 2D positional encoder for SATRN, see
`SATRN <https://arxiv.org/abs/1910.04396>`_
Modified from https://github.com/Media-Smart/vedastr
Licensed under the Apache License, Version 2.0 (the "License");
Args:
d_hid (int): Dimensions of hidden layer.
n_height (int): Max height of the 2D feature output.
n_width (int): Max width of the 2D feature output.
dropout (int): Size of hidden layers of the model.
"""
def __init__(self,
d_hid=512,
n_height=100,
n_width=100,
dropout=0.1):
super().__init__()
h_position_encoder = self._get_sinusoid_encoding_table(n_height, d_hid)
h_position_encoder = h_position_encoder.transpose(0, 1)
h_position_encoder = h_position_encoder.view(1, d_hid, n_height, 1)
w_position_encoder = self._get_sinusoid_encoding_table(n_width, d_hid)
w_position_encoder = w_position_encoder.transpose(0, 1)
w_position_encoder = w_position_encoder.view(1, d_hid, 1, n_width)
self.register_buffer('h_position_encoder', h_position_encoder)
self.register_buffer('w_position_encoder', w_position_encoder)
self.h_scale = self.scale_factor_generate(d_hid)
self.w_scale = self.scale_factor_generate(d_hid)
self.pool = nn.AdaptiveAvgPool2d(1)
self.dropout = nn.Dropout(p=dropout)
def _get_sinusoid_encoding_table(self, n_position, d_hid):
"""Sinusoid position encoding table."""
denominator = torch.Tensor([
1.0 / np.power(10000, 2 * (hid_j // 2) / d_hid)
for hid_j in range(d_hid)
])
denominator = denominator.view(1, -1)
pos_tensor = torch.arange(n_position).unsqueeze(-1).float()
sinusoid_table = pos_tensor * denominator
sinusoid_table[:, 0::2] = torch.sin(sinusoid_table[:, 0::2])
sinusoid_table[:, 1::2] = torch.cos(sinusoid_table[:, 1::2])
return sinusoid_table
def scale_factor_generate(self, d_hid):
scale_factor = nn.Sequential(
nn.Conv2d(d_hid, d_hid, kernel_size=1), nn.ReLU(inplace=True),
nn.Conv2d(d_hid, d_hid, kernel_size=1), nn.Sigmoid())
return scale_factor
def forward(self, x):
b, c, h, w = x.size()
avg_pool = self.pool(x)
h_pos_encoding = \
self.h_scale(avg_pool) * self.h_position_encoder[:, :, :h, :]
w_pos_encoding = \
self.w_scale(avg_pool) * self.w_position_encoder[:, :, :, :w]
out = x + h_pos_encoding + w_pos_encoding
out = self.dropout(out)
return out
class SatrnEncoder(nn.Module):
"""Implement encoder for SATRN, see `SATRN.
<https://arxiv.org/abs/1910.04396>`_.
Args:
n_layers (int): Number of attention layers.
n_head (int): Number of parallel attention heads.
d_k (int): Dimension of the key vector.
d_v (int): Dimension of the value vector.
d_model (int): Dimension :math:`D_m` of the input from previous model.
n_position (int): Length of the positional encoding vector. Must be
greater than ``max_seq_len``.
d_inner (int): Hidden dimension of feedforward layers.
dropout (float): Dropout rate.
init_cfg (dict or list[dict], optional): Initialization configs.
"""
def __init__(self,
n_layers=12,
n_head=8,
d_k=64,
d_v=64,
d_model=512,
n_position=100,
d_inner=256,
dropout=0.1,
**kwargs):
super().__init__()
self.d_model = d_model
self.position_enc = Adaptive2DPositionalEncoding(
d_hid=d_model,
n_height=n_position,
n_width=n_position,
dropout=dropout)
self.layer_stack = nn.ModuleList([
SatrnEncoderLayer(
d_model, d_inner, n_head, d_k, d_v, dropout=dropout)
for _ in range(n_layers)
])
self.layer_norm = nn.LayerNorm(d_model)
def forward(self, feat):
"""
Args:
feat (Tensor): Feature tensor of shape :math:`(N, D_m, H, W)`.
Returns:
Tensor: A tensor of shape :math:`(N, T, D_m)`.
"""
feat = feat + self.position_enc(feat)
n, c, h, w = feat.size()
mask = feat.new_zeros((n, h, w))
mask[:,:,:w] = 1
mask = mask.view(n, h * w)
feat = feat.view(n, c, h * w)
output = feat.permute(0, 2, 1).contiguous()
for enc_layer in self.layer_stack:
output = enc_layer(output, h, w, mask)
output = self.layer_norm(output)
return output

View File

@ -1,57 +0,0 @@
import os
import shutil
import colorama
import hashlib
def sha256sum(filename):
h = hashlib.sha256()
b = bytearray(128*1024)
mv = memoryview(b)
with open(filename, 'rb', buffering=0) as f:
for n in iter(lambda : f.readinto(mv), 0):
h.update(mv[:n])
return h.hexdigest()
online_model_factory = {
'satrn-lite-general-pretrain-20230106': {
'url': 'https://github.com/moewiee/satrn-model-factory/raw/main/jxqhbem4to.pth',
'hash': 'b0069a72bf6fc080ad5d431d5ce650c3bfbab535141adef1631fce331cb1471c'},
'satrn-lite-captcha-finetune-20230108': {
'url': 'https://github.com/moewiee/satrn-model-factory/raw/main/l27vitogmc.pth',
'hash': 'efcbcf2955b6b21125b073f83828d2719e908c7303b0d9901e94be5a8efba916'},
'satrn-lite-handwritten-finetune-20230108': {
'url': 'https://github.com/moewiee/satrn-model-factory/raw/main/lj9gkwelns.pth',
'hash': 'bccd8e985b131fcd4701114af5ceaef098f2eac50654bbb1d828e7f829e711dd'},
}
__hub_available_versions__ = online_model_factory.keys()
def _get(version):
use_online = version in __hub_available_versions__
if not use_online and not os.path.exists(version):
raise ValueError(f'Model version {version} not found online and not found local.')
hub_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), 'hub')
if not os.path.exists(hub_path):
os.makedirs(hub_path)
if use_online:
version_url = online_model_factory[version]['url']
file_path = os.path.join(hub_path, os.path.basename(version_url))
else:
file_path = os.path.join(hub_path, os.path.basename(version))
if not os.path.exists(file_path):
if use_online:
os.system(f'wget -O {file_path} {version_url}')
assert os.path.exists(file_path), \
colorama.Fore.RED + 'wget failed while trying to retrieve from hub.' + colorama.Style.RESET_ALL
downloaded_hash = sha256sum(file_path)
if downloaded_hash != online_model_factory[version]['hash']:
os.remove(file_path)
raise ValueError(colorama.Fore.RED + 'sha256 hash doesnt match for version retrieved from hub.' + colorama.Style.RESET_ALL)
else:
shutil.copy2(version, file_path)
return file_path

View File

@ -1,78 +0,0 @@
import torch
import torch.nn as nn
import numpy as np
from collections import abc
def cast_tensor_type(inputs, src_type, dst_type):
"""Recursively convert Tensor in inputs from src_type to dst_type.
Args:
inputs: Inputs that to be casted.
src_type (torch.dtype): Source type..
dst_type (torch.dtype): Destination type.
Returns:
The same type with inputs, but all contained Tensors have been cast.
"""
if isinstance(inputs, nn.Module):
return inputs
elif isinstance(inputs, torch.Tensor):
return inputs.to(dst_type)
elif isinstance(inputs, str):
return inputs
elif isinstance(inputs, np.ndarray):
return inputs
elif isinstance(inputs, abc.Mapping):
return type(inputs)({
k: cast_tensor_type(v, src_type, dst_type)
for k, v in inputs.items()
})
elif isinstance(inputs, abc.Iterable):
return type(inputs)(
cast_tensor_type(item, src_type, dst_type) for item in inputs)
else:
return inputs
def patch_forward_method(func, src_type, dst_type, convert_output=True):
"""Patch the forward method of a module.
Args:
func (callable): The original forward method.
src_type (torch.dtype): Type of input arguments to be converted from.
dst_type (torch.dtype): Type of input arguments to be converted to.
convert_output (bool): Whether to convert the output back to src_type.
Returns:
callable: The patched forward method.
"""
def new_forward(*args, **kwargs):
output = func(*cast_tensor_type(args, src_type, dst_type),
**cast_tensor_type(kwargs, src_type, dst_type))
if convert_output:
output = cast_tensor_type(output, dst_type, src_type)
return output
return new_forward
def patch_norm_fp32(module):
"""Recursively convert normalization layers from FP16 to FP32.
Args:
module (nn.Module): The modules to be converted in FP16.
Returns:
nn.Module: The converted module, the normalization layers have been
converted to FP32.
"""
if isinstance(module, (nn.modules.batchnorm._BatchNorm, nn.GroupNorm)):
module.float()
if isinstance(module, nn.GroupNorm) or torch.__version__ < '1.3':
module.forward = patch_forward_method(module.forward, torch.half,
torch.float)
for child in module.children():
patch_norm_fp32(child)
return module

View File

@ -1,33 +0,0 @@
import torchvision.transforms.functional as TF
import cv2
import torch
class DataPipelineSATRN:
def __init__(self,
resize_height,
resize_width,
norm_mean,
norm_std,
device='cpu'):
self.resize_height = resize_height
self.resize_width = resize_width
self.norm_mean = norm_mean
self.norm_std = norm_std
self.device = device
def __call__(self, imgs):
if not isinstance(imgs, list):
imgs = [imgs]
datas = []
for img in imgs:
if isinstance(img, str):
img = cv2.imread(img)
data = torch.from_numpy(cv2.resize(img, (self.resize_width, self.resize_height), interpolation=cv2.INTER_LINEAR))
datas.append(data)
data = torch.stack(datas, dim=0)
data = data.to(self.device)
data = data.float().div_(255.).permute((0,3,1,2))
TF.normalize(data, mean=self.norm_mean, std=self.norm_std, inplace=True)
return data.half() if self.device != 'cpu' else data

View File

@ -1,12 +0,0 @@
import numpy as np
from sdsvtr import StandaloneSATRNRunner
runner = StandaloneSATRNRunner(version='satrn-lite-general-pretrain-20230106', return_confident=False, device='cuda:1')
dummy_list = [np.ndarray((32,128,3)) for _ in range(100)]
bs = 32
all_results = []
while len(dummy_list) > 0:
dummy_batch = dummy_list[:bs]
dummy_list = dummy_list[bs:]
all_results += runner(dummy_batch)

View File

@ -1,82 +0,0 @@
addict==2.4.0
asttokens==2.2.1
autopep8==1.6.0
backcall==0.2.0
backports.functools-lru-cache==1.6.4
brotlipy==0.7.0
certifi==2022.12.7
cffi==1.15.1
charset-normalizer==2.0.4
click==8.1.3
colorama==0.4.6
cryptography==39.0.1
debugpy==1.5.1
decorator==5.1.1
docopt==0.6.2
entrypoints==0.4
executing==1.2.0
flit_core==3.6.0
idna==3.4
importlib-metadata==6.0.0
ipykernel==6.15.0
ipython==8.11.0
jedi==0.18.2
jupyter-client==7.0.6
jupyter_core==4.12.0
Markdown==3.4.1
markdown-it-py==2.2.0
matplotlib-inline==0.1.6
mdurl==0.1.2
mkl-fft==1.3.1
mkl-random==1.2.2
mkl-service==2.4.0
mmcv-full==1.7.1
model-index==0.1.11
nest-asyncio==1.5.6
numpy==1.23.5
opencv-python==4.7.0.72
openmim==0.3.6
ordered-set==4.1.0
packaging==23.0
pandas==1.5.3
parso==0.8.3
pexpect==4.8.0
pickleshare==0.7.5
Pillow==9.4.0
pip==22.3.1
pipdeptree==2.5.2
prompt-toolkit==3.0.38
psutil==5.9.0
ptyprocess==0.7.0
pure-eval==0.2.2
pycodestyle==2.10.0
pycparser==2.21
Pygments==2.14.0
pyOpenSSL==23.0.0
PySocks==1.7.1
python-dateutil==2.8.2
pytz==2022.7.1
PyYAML==6.0
pyzmq==19.0.2
requests==2.28.1
rich==13.3.1
sdsvtd==0.1.1
sdsvtr==0.0.5
setuptools==65.6.3
Shapely==1.8.4
six==1.16.0
stack-data==0.6.2
tabulate==0.9.0
toml==0.10.2
torch==1.13.1
torchvision==0.14.1
tornado==6.1
tqdm==4.65.0
traitlets==5.9.0
typing_extensions==4.4.0
urllib3==1.26.14
wcwidth==0.2.6
wheel==0.38.4
yapf==0.32.0
yarg==0.1.9
zipp==3.15.0

View File

@ -1,200 +0,0 @@
"""
see scripts/run_ocr.sh to run
"""
# from pathlib import Path # add parent path to run debugger
# import sys
# FILE = Path(__file__).absolute()
# sys.path.append(FILE.parents[2].as_posix())
from src.utils import construct_file_path, ImageReader
from src.dto import Line
from src.ocr import OcrEngine
import argparse
import tqdm
import pandas as pd
from pathlib import Path
import json
import os
import numpy as np
from typing import Union, Tuple, List, Optional
from collections import defaultdict
current_dir = os.getcwd()
def get_args():
parser = argparse.ArgumentParser()
# parser image
parser.add_argument(
"--image",
type=str,
required=True,
help="path to input image/directory/csv file",
)
parser.add_argument(
"--save_dir", type=str, required=True, help="path to save directory"
)
parser.add_argument(
"--include", type=str, nargs="+", default=[], help="files/folders to include"
)
parser.add_argument(
"--exclude", type=str, nargs="+", default=[], help="files/folders to exclude"
)
parser.add_argument(
"--base_dir",
type=str,
required=False,
default=current_dir,
help="used when --image and --save_dir are relative paths to a base directory, default to current directory",
)
parser.add_argument(
"--export_csv",
type=str,
required=False,
default="",
help="used when --image is a directory. If set, a csv file contains image_path, ocr_path and label will be exported to save_dir.",
)
parser.add_argument(
"--export_img",
type=bool,
required=False,
default=False,
help="whether to save the visualize img",
)
parser.add_argument("--ocr_kwargs", type=str, required=False, default="")
opt = parser.parse_args()
return opt
def load_engine(opt) -> OcrEngine:
print("[INFO] Loading engine...")
kw = json.loads(opt.ocr_kwargs) if opt.ocr_kwargs else {}
engine = OcrEngine(**kw)
print("[INFO] Engine loaded")
return engine
def convert_relative_path_to_positive_path(tgt_dir: Path, base_dir: Path) -> Path:
return tgt_dir if tgt_dir.is_absolute() else base_dir.joinpath(tgt_dir)
def get_paths_from_opt(opt) -> Tuple[Path, Path]:
# BC\ kiem\ tra\ y\ te -> BC kiem tra y te
img_path = opt.image.replace("\\ ", " ").strip()
save_dir = opt.save_dir.replace("\\ ", " ").strip()
base_dir = opt.base_dir.replace("\\ ", " ").strip()
input_image = convert_relative_path_to_positive_path(Path(img_path), Path(base_dir))
save_dir = convert_relative_path_to_positive_path(Path(save_dir), Path(base_dir))
if not save_dir.exists():
save_dir.mkdir()
print("[INFO]: Creating folder ", save_dir)
return input_image, save_dir
def process_img(
img: Union[str, np.ndarray],
save_dir_or_path: str,
engine: OcrEngine,
export_img: bool,
save_path_deskew: Optional[str] = None,
) -> None:
save_dir_or_path = Path(save_dir_or_path)
if isinstance(img, np.ndarray):
if save_dir_or_path.is_dir():
raise ValueError("numpy array input require a save path, not a save dir")
page = engine(img)
save_path = (
str(save_dir_or_path.joinpath(Path(img).stem + ".txt"))
if save_dir_or_path.is_dir()
else str(save_dir_or_path)
)
page.write_to_file("word", save_path)
if export_img:
page.save_img(
save_path.replace(".txt", ".jpg"),
is_vnese=True,
save_path_deskew=save_path_deskew,
)
def process_dir(
dir_path: str,
save_dir: str,
engine: OcrEngine,
export_img: bool,
lexcludes: List[str] = [],
lincludes: List[str] = [],
ddata=defaultdict(list),
) -> None:
pdir_path = Path(dir_path)
print(pdir_path)
# save_dir_sub = Path(construct_file_path(save_dir, dir_path, ext=""))
psave_dir = Path(save_dir)
psave_dir.mkdir(exist_ok=True)
for img_path in (pbar := tqdm.tqdm(pdir_path.iterdir())):
pbar.set_description(f"Processing {pdir_path}")
if (lincludes and img_path.name not in lincludes) or (
img_path.name in lexcludes
):
continue # only process desired files/foders
if img_path.is_dir():
psave_dir_sub = psave_dir.joinpath(img_path.stem)
process_dir(img_path, str(psave_dir_sub), engine, ddata)
elif img_path.suffix.lower() in ImageReader.supported_ext:
simg_path = str(img_path)
# try:
img = (
ImageReader.read(simg_path)
if img_path.suffix != ".pdf"
else ImageReader.read(simg_path)[0]
)
save_path = str(Path(psave_dir).joinpath(img_path.stem + ".txt"))
save_path_deskew = str(
Path(psave_dir).joinpath(img_path.stem + "_deskewed.jpg")
)
process_img(img, save_path, engine, export_img, save_path_deskew)
# except Exception as e:
# print('[ERROR]: ', e, ' at ', simg_path)
# continue
ddata["img_path"].append(simg_path)
ddata["ocr_path"].append(save_path)
if Path(save_path_deskew).exists():
ddata["save_path_deskew"].append(save_path)
ddata["label"].append(pdir_path.stem)
# ddata.update({"img_path": img_path, "save_path": save_path, "label": dir_path.stem})
return ddata
def process_csv(csv_path: str, engine: OcrEngine) -> None:
df = pd.read_csv(csv_path)
if not "image_path" in df.columns or not "ocr_path" in df.columns:
raise AssertionError("Cannot fing image_path in df headers")
for row in df.iterrows():
process_img(row.image_path, row.ocr_path, engine)
if __name__ == "__main__":
opt = get_args()
engine = load_engine(opt)
print("[INFO]: OCR engine settings:", engine.settings)
img, save_dir = get_paths_from_opt(opt)
lskip_dir = []
if img.is_dir():
ddata = process_dir(
img, save_dir, engine, opt.export_img, opt.exclude, opt.include
)
if opt.export_csv:
pd.DataFrame.from_dict(ddata).to_csv(
Path(save_dir).joinpath(opt.export_csv)
)
elif img.suffix in ImageReader.supported_ext:
process_img(str(img), save_dir, engine, opt.export_img)
elif img.suffix == ".csv":
print(
"[WARNING]: Running with csv file will ignore the save_dir argument. Instead, the ocr_path in the csv would be used"
)
process_csv(img, engine)
else:
raise NotImplementedError("[ERROR]: Unsupported file {}".format(img))

View File

@ -1,9 +0,0 @@
export CUDA_VISIBLE_DEVICES=1
# export PATH=/usr/local/cuda-11.6/bin${PATH:+:${PATH}}
# export LD_LIBRARY_PATH=/usr/local/cuda-11.6/lib64\ {LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
# export CUDA_HOME=/usr/local/cuda-11.6
# export PATH=/usr/local/cuda-11.6/bin:$PATH
# export CPATH=/usr/local/cuda-11.6/include:$CPATH
# export LIBRARY_PATH=/usr/local/cuda-11.6/lib64:$LIBRARY_PATH
# export LD_LIBRARY_PATH=/usr/local/cuda-11.6/lib64:/usr/local/cuda-11.6/extras/CUPTI/lib64:$LD_LIBRARY_PATH
python test/test_deskew_dir.py

View File

@ -1,49 +0,0 @@
#bash scripts/run_ocr.sh -i /mnt/hdd2T/AICR/Projects/2023/FWD/Forms/PDFs/ -o /mnt/ssd1T/hungbnt/DocumentClassification/results/ocr -e out.csv -k "{\"device\":\"cuda:1\"}" -p True -n Passport 'So\ HK'
#bash scripts/run_ocr.sh -i '/mnt/hdd2T/AICR/Projects/2023/FWD/Forms/PDFs/So\ HK' -o /mnt/ssd1T/hungbnt/DocumentClassification/results/ocr -e out.csv -k "{\"device\":\"cuda:1\"}" -p True
#-n and -x do not accept multiple argument currently
# bash scripts/run_ocr.sh -i /mnt/hdd4T/OCR/hoangdc/End_to_end/ICDAR2013/data/images_receipt_5images/ -o visualize/ -e out.csv -k "{\"device\":\"cuda:1\"}" -p True
export PYTHONWARNINGS="ignore"
while getopts i:o:b:e:p:k:n:x: flag
do
case "${flag}" in
i) img=${OPTARG};;
o) out_dir=${OPTARG};;
b) base_dir=${OPTARG};;
e) export_csv=${OPTARG};;
p) export_img=${OPTARG};;
k) ocr_kwargs=${OPTARG};;
n) include=("${OPTARG[@]}");;
x) exclude=("${OPTARG[@]}");;
esac
done
cmd="python run.py \
--image $img \
--save_dir $out_dir \
--export_csv $export_csv \
--export_img $export_img \
--ocr_kwargs $ocr_kwargs"
if [ ${#include[@]} -gt 0 ]; then
cmd+=" --include"
for item in "${include[@]}"; do
cmd+=" $item"
done
fi
if [ ${#exclude[@]} -gt 0 ]; then
cmd+=" --exclude"
for item in "${exclude[@]}"; do
cmd+=" $item"
done
fi
echo $cmd
exec $cmd

View File

@ -1,35 +0,0 @@
device: &device cuda:0
max_img_size: [1920,1920] #text det default size: 1280x1280 #[] = originla size, TODO: fix the deskew code to resize the image only for detecting the angle, we want to feed the original size image to the text detection pipeline so that the bounding boxes would be mapped back to the original size
extend_bbox: [0, 0.0, 0.0, 0.0] # left, top, right, bottom
batch_size: 1 #1 means batch_mode = False
detector:
# version: /mnt/hdd2T/datnt/datnt_from_ssd1T/mmdetection/wild_receipt_finetune_weights_c_lite.pth
version: /mnt/hdd4T/OCR/datnt/mmdetection/logs/textdet-baseline-Oct04-wildreceiptv4-sdsapv1-mcocr-ssreceipt/epoch_100_params.pth
auto_rotate: True
rotator_version: /mnt/hdd2T/datnt/datnt_from_ssd1T/mmdetection/logs/textdet-with-rotate-20230317/best_bbox_mAP_epoch_30_lite.pth
device: *device
recognizer:
version: satrn-lite-general-pretrain-20230106
max_seq_len_overwrite: 24 #default = 12
return_confident: True
device: *device
#extend the bbox to avoid losing accent mark in vietnames, if using ocr for only english, disable it
deskew:
enable: True
text_detector:
config: /mnt/hdd4T/OCR/tuanlv/01-BasicOCR/ocr-engine-deskew/externals/sdsv_dewarp/config/det.yaml
weight: /mnt/hdd4T/OCR/tuanlv/01-BasicOCR/ocr-engine-deskew/externals/sdsv_dewarp/weights/ch_PP-OCRv3_det_infer
text_cls:
config: /mnt/hdd4T/OCR/tuanlv/01-BasicOCR/ocr-engine-deskew/externals/sdsv_dewarp/config/cls.yaml
weight: /mnt/hdd4T/OCR/tuanlv/01-BasicOCR/ocr-engine-deskew/externals/sdsv_dewarp/weights/ch_ppocr_mobile_v2.0_cls_infer
device: *device
words_to_lines:
gradient: 0.6
max_x_dist: 20
max_running_y_shift_degree: 10 #degrees
y_overlap_threshold: 0.5
word_formation_mode: line

View File

@ -1,534 +0,0 @@
import numpy as np
from typing import Optional, List, Union
import cv2
from PIL import Image
from pathlib import Path
from .utils import visualize_bbox_and_label
class Box:
def __init__(
self, x1: int, y1: int, x2: int, y2: int, conf: float = -1.0, label: str = ""
):
self._x1 = x1
self._y1 = y1
self._x2 = x2
self._y2 = y2
self._conf = conf
self._label = label
def __repr__(self) -> str:
return str(self.bbox)
def __str__(self) -> str:
return str(self.bbox)
def get(self, return_confidence=False) -> Union[list[int], list[Union[float, int]]]:
return self.bbox if not return_confidence else self.xyxyc
def __getitem__(self, key):
return self.bbox[key]
@property
def width(self):
return max(self._x2 - self._x1, -1)
@property
def height(self):
return max(self._y2 - self._y1, -1)
@property
def bbox(self) -> list[int]:
return [self._x1, self._y1, self._x2, self._y2]
@bbox.setter
def bbox(self, bbox_: list[int]):
self._x1, self._y1, self._x2, self._y2 = bbox_
@property
def xyxyc(self) -> list[Union[float, int]]:
return [self._x1, self._y1, self._x2, self._y2, self._conf]
@staticmethod
def normalize_bbox(bbox: list[int]) -> list[int]:
return [int(b) for b in bbox]
def to_int(self):
self._x1, self._y1, self._x2, self._y2 = self.normalize_bbox(
[self._x1, self._y1, self._x2, self._y2]
)
return self
@staticmethod
def clamp_bbox_by_img_wh(bbox: list, width: int, height: int) -> list[int]:
x1, y1, x2, y2 = bbox
x1 = min(max(0, x1), width)
x2 = min(max(0, x2), width)
y1 = min(max(0, y1), height)
y2 = min(max(0, y2), height)
return [x1, y1, x2, y2]
def clamp_by_img_wh(self, width: int, height: int):
self._x1, self._y1, self._x2, self._y2 = self.clamp_bbox_by_img_wh(
[self._x1, self._y1, self._x2, self._y2], width, height
)
return self
@staticmethod
def extend_bbox(bbox: list, margin: list): # -> Self (python3.11)
margin_l, margin_t, margin_r, margin_b = margin
l, t, r, b = bbox # left, top, right, bottom
t = t - (b - t) * margin_t
b = b + (b - t) * margin_b
l = l - (r - l) * margin_l
r = r + (r - l) * margin_r
return [l, t, r, b]
def get_extend_bbox(self, margin: list):
extended_bbox = self.extend_bbox(self.bbox, margin)
return Box(*extended_bbox, label=self._label)
@staticmethod
def bbox_is_valid(bbox: list[int]) -> bool:
if bbox == [-1, -1, -1, -1]:
raise ValueError("Empty bounding box found")
l, t, r, b = bbox # left, top, right, bottom
return True if (b - t) * (r - l) > 0 else False
def is_valid(self) -> bool:
return self.bbox_is_valid(self.bbox)
@staticmethod
def crop_img_by_bbox(img: np.ndarray, bbox: list) -> np.ndarray:
l, t, r, b = bbox
return img[t:b, l:r]
def crop_img(self, img: np.ndarray) -> np.ndarray:
return self.crop_img_by_bbox(img, self.bbox)
class Word:
def __init__(
self,
image=None,
text="",
conf_cls=-1.0,
bbox_obj: Box = Box(-1, -1, -1, -1),
conf_detect=-1.0,
kie_label="",
):
# self.type = "word"
self._text = text
self._image = image
self._conf_det = conf_detect
self._conf_cls = conf_cls
# [left, top,right,bot] coordinate of top-left and bottom-right point
self._bbox_obj = bbox_obj
# self.word_id = 0 # id of word
# self.word_group_id = 0 # id of word_group which instance belongs to
# self.line_id = 0 # id of line which instance belongs to
# self.paragraph_id = 0 # id of line which instance belongs to
self._kie_label = kie_label
@property
def bbox(self) -> list[int]:
return self._bbox_obj.bbox
@property
def text(self) -> str:
return self._text
@property
def height(self):
return self._bbox_obj.height
@property
def width(self):
return self._bbox_obj.width
def __repr__(self) -> str:
return self._text
def __str__(self) -> str:
return self._text
def is_valid(self) -> bool:
return self._bbox_obj.is_valid()
# def is_special_word(self):
# if not self._text:
# raise ValueError("Cannot validatie size of empty bounding box")
# # if len(text) > 7:
# # return True
# if len(self._text) >= 7:
# no_digits = sum(c.isdigit() for c in text)
# return no_digits / len(text) >= 0.3
# return False
class WordGroup:
def __init__(
self,
list_words: List[Word] = list(),
text: str = "",
boundingbox: Box = Box(-1, -1, -1, -1),
conf_cls: float = -1,
conf_det: float = -1,
):
# self.type = "word_group"
self._list_words = list_words # dict of word instances
# self.word_group_id = 0 # word group id
# self.line_id = 0 # id of line which instance belongs to
# self.paragraph_id = 0 # id of paragraph which instance belongs to
self._text = text
self._bbox_obj = boundingbox
self._kie_label = ""
self._conf_cls = conf_cls
self._conf_det = conf_det
@property
def bbox(self) -> list[int]:
return self._bbox_obj.bbox
@property
def text(self) -> str:
return self._text
@property
def list_words(self) -> list[Word]:
return self._list_words
def __repr__(self) -> str:
return self._text
def __str__(self) -> str:
return self._text
# def add_word(self, word: Word): # add a word instance to the word_group
# if word._text != "✪":
# for w in self._list_words:
# if word.word_id == w.word_id:
# print("Word id collision")
# return False
# word.word_group_id = self.word_group_id #
# word.line_id = self.line_id
# word.paragraph_id = self.paragraph_id
# self._list_words.append(word)
# self._text += " " + word._text
# if self.bbox_obj == [-1, -1, -1, -1]:
# self.bbox_obj = word._bbox_obj
# else:
# self.bbox_obj = [
# min(self.bbox_obj[0], word._bbox_obj[0]),
# min(self.bbox_obj[1], word._bbox_obj[1]),
# max(self.bbox_obj[2], word._bbox_obj[2]),
# max(self.bbox_obj[3], word._bbox_obj[3]),
# ]
# return True
# else:
# return False
# def update_word_group_id(self, new_word_group_id):
# self.word_group_id = new_word_group_id
# for i in range(len(self._list_words)):
# self._list_words[i].word_group_id = new_word_group_id
# def update_kie_label(self):
# list_kie_label = [word._kie_label for word in self._list_words]
# dict_kie = dict()
# for label in list_kie_label:
# if label not in dict_kie:
# dict_kie[label] = 1
# else:
# dict_kie[label] += 1
# total = len(list(dict_kie.values()))
# max_value = max(list(dict_kie.values()))
# list_keys = list(dict_kie.keys())
# list_values = list(dict_kie.values())
# self.kie_label = list_keys[list_values.index(max_value)]
# def update_text(self): # update text after changing positions of words in list word
# text = ""
# for word in self._list_words:
# text += " " + word._text
# self._text = text
class Line:
def __init__(
self,
list_word_groups: List[WordGroup] = [],
text: str = "",
boundingbox: Box = Box(-1, -1, -1, -1),
conf_cls: float = -1,
conf_det: float = -1,
):
# self.type = "line"
self._list_word_groups = (
list_word_groups # list of Word_group instances in the line
)
# self.line_id = 0 # id of line in the paragraph
# self.paragraph_id = 0 # id of paragraph which instance belongs to
self._text = text
self._bbox_obj = boundingbox
self._conf_cls = conf_cls
self._conf_det = conf_det
@property
def bbox(self) -> list[int]:
return self._bbox_obj.bbox
@property
def text(self) -> str:
return self._text
@property
def list_word_groups(self) -> List[WordGroup]:
return self._list_word_groups
@property
def list_words(self) -> list[Word]:
return [
word
for word_group in self._list_word_groups
for word in word_group.list_words
]
def __repr__(self) -> str:
return self._text
def __str__(self) -> str:
return self._text
# def add_group(self, word_group: WordGroup): # add a word_group instance
# if word_group._list_words is not None:
# for wg in self.list_word_groups:
# if word_group.word_group_id == wg.word_group_id:
# print("Word_group id collision")
# return False
# self.list_word_groups.append(word_group)
# self.text += word_group._text
# word_group.paragraph_id = self.paragraph_id
# word_group.line_id = self.line_id
# for i in range(len(word_group._list_words)):
# word_group._list_words[
# i
# ].paragraph_id = self.paragraph_id # set paragraph_id for word
# word_group._list_words[i].line_id = self.line_id # set line_id for word
# return True
# return False
# def update_line_id(self, new_line_id):
# self.line_id = new_line_id
# for i in range(len(self.list_word_groups)):
# self.list_word_groups[i].line_id = new_line_id
# for j in range(len(self.list_word_groups[i]._list_words)):
# self.list_word_groups[i]._list_words[j].line_id = new_line_id
# def merge_word(self, word): # word can be a Word instance or a Word_group instance
# if word.text != "✪":
# if self.boundingbox == [-1, -1, -1, -1]:
# self.boundingbox = word.boundingbox
# else:
# self.boundingbox = [
# min(self.boundingbox[0], word.boundingbox[0]),
# min(self.boundingbox[1], word.boundingbox[1]),
# max(self.boundingbox[2], word.boundingbox[2]),
# max(self.boundingbox[3], word.boundingbox[3]),
# ]
# self.list_word_groups.append(word)
# self.text += " " + word.text
# return True
# return False
# def __cal_ratio(self, top1, bottom1, top2, bottom2):
# sorted_vals = sorted([top1, bottom1, top2, bottom2])
# intersection = sorted_vals[2] - sorted_vals[1]
# min_height = min(bottom1 - top1, bottom2 - top2)
# if min_height == 0:
# return -1
# ratio = intersection / min_height
# return ratio
# def __cal_ratio_height(self, top1, bottom1, top2, bottom2):
# height1, height2 = top1 - bottom1, top2 - bottom2
# ratio_height = float(max(height1, height2)) / float(min(height1, height2))
# return ratio_height
# def in_same_line(self, input_line, thresh=0.7):
# # calculate iou in vertical direction
# _, top1, _, bottom1 = self.boundingbox
# _, top2, _, bottom2 = input_line.boundingbox
# ratio = self.__cal_ratio(top1, bottom1, top2, bottom2)
# ratio_height = self.__cal_ratio_height(top1, bottom1, top2, bottom2)
# if (
# (top2 <= top1 <= bottom2) or (top1 <= top2 <= bottom1)
# and ratio >= thresh
# and (ratio_height < 2)
# ):
# return True
# return False
# class Paragraph:
# def __init__(self, id=0, lines=None):
# self.list_lines = lines if lines is not None else [] # list of all lines in the paragraph
# self.paragraph_id = id # index of paragraph in the ist of paragraph
# self.text = ""
# self.boundingbox = [-1, -1, -1, -1]
# @property
# def bbox(self):
# return self.boundingbox
# def __repr__(self) -> str:
# return self.text
# def __str__(self) -> str:
# return self.text
# def add_line(self, line: Line): # add a line instance
# if line.list_word_groups is not None:
# for l in self.list_lines:
# if line.line_id == l.line_id:
# print("Line id collision")
# return False
# for i in range(len(line.list_word_groups)):
# line.list_word_groups[
# i
# ].paragraph_id = (
# self.paragraph_id
# ) # set paragraph id for every word group in line
# for j in range(len(line.list_word_groups[i]._list_words)):
# line.list_word_groups[i]._list_words[
# j
# ].paragraph_id = (
# self.paragraph_id
# ) # set paragraph id for every word in word groups
# line.paragraph_id = self.paragraph_id # set paragraph id for line
# self.list_lines.append(line) # add line to paragraph
# self.text += " " + line.text
# return True
# else:
# return False
# def update_paragraph_id(
# self, new_paragraph_id
# ): # update new paragraph_id for all lines, word_groups, words inside paragraph
# self.paragraph_id = new_paragraph_id
# for i in range(len(self.list_lines)):
# self.list_lines[
# i
# ].paragraph_id = new_paragraph_id # set new paragraph_id for line
# for j in range(len(self.list_lines[i].list_word_groups)):
# self.list_lines[i].list_word_groups[
# j
# ].paragraph_id = new_paragraph_id # set new paragraph_id for word_group
# for k in range(len(self.list_lines[i].list_word_groups[j].list_words)):
# self.list_lines[i].list_word_groups[j].list_words[
# k
# ].paragraph_id = new_paragraph_id # set new paragraph id for word
# return True
class Page:
def __init__(
self,
word_segments: Union[List[WordGroup], List[Line]],
image: np.ndarray,
deskewed_image: Optional[np.ndarray] = None,
) -> None:
self._word_segments = word_segments
self._image = image
self._deskewed_image = deskewed_image
self._drawed_image: Optional[np.ndarray] = None
@property
def word_segments(self):
return self._word_segments
@property
def list_words(self) -> list[Word]:
return [
word
for word_segment in self._word_segments
for word in word_segment.list_words
]
@property
def image(self):
return self._image
@property
def PIL_image(self):
return Image.fromarray(self._image)
@property
def drawed_image(self):
return self._drawed_image
@property
def deskewed_image(self):
return self._deskewed_image
def visualize_bbox_and_label(self, **kwargs: dict):
if self._drawed_image is not None:
return self._drawed_image
bboxes = list()
texts = list()
for word in self.list_words:
bboxes.append([int(float(b)) for b in word.bbox])
texts.append(word._text)
img = visualize_bbox_and_label(
self._deskewed_image if self._deskewed_image is not None else self._image,
bboxes,
texts,
**kwargs
)
self._drawed_image = img
return self._drawed_image
def save_img(self, save_path: str, **kwargs: dict) -> None:
save_path_deskew = kwargs.pop("save_path_deskew", Path(save_path).with_stem(Path(save_path).stem + "_deskewed").as_posix())
if self._deskewed_image is not None:
# save_path_deskew: str = kwargs.pop("save_path_deskew", Path(save_path).with_stem(Path(save_path).stem + "_deskewed").as_posix()) # type: ignore
cv2.imwrite(save_path_deskew, self._deskewed_image)
img = self.visualize_bbox_and_label(**kwargs)
cv2.imwrite(save_path, img)
def write_to_file(self, mode: str, save_path: str) -> None:
f = open(save_path, "w+", encoding="utf-8")
for word_segment in self._word_segments:
if mode == "segment":
xmin, ymin, xmax, ymax = word_segment.bbox
f.write(
"{}\t{}\t{}\t{}\t{}\n".format(
xmin, ymin, xmax, ymax, word_segment._text
)
)
elif mode == "word":
for word in word_segment.list_words:
# xmin, ymin, xmax, ymax = word.bbox
xmin, ymin, xmax, ymax = [int(float(b)) for b in word.bbox]
f.write(
"{}\t{}\t{}\t{}\t{}\n".format(
xmin, ymin, xmax, ymax, word._text
)
)
else:
raise NotImplementedError("Unknown mode: {}".format(mode))
f.close()
class Document:
def __init__(self, lpages: List[Page]) -> None:
self.lpages = lpages

View File

@ -1,258 +0,0 @@
from typing import Union, overload, List, Optional, Tuple
from PIL import Image
import torch
import numpy as np
import yaml
from pathlib import Path
import mmcv
from sdsvtd import StandaloneYOLOXRunner
from sdsvtr import StandaloneSATRNRunner
from sdsv_dewarp.api import AlignImage
from .utils import ImageReader, chunks, Timer, post_process_recog # rotate_bbox
# from .utils import jdeskew as deskew
# from externals.deskew.sdsv_dewarp import pdeskew as deskew
# from .utils import deskew
from .dto import Word, Line, Page, Document, Box, WordGroup
# from .word_formation import words_to_lines as words_to_lines
# from .word_formation import wo rds_to_lines_mmocr as words_to_lines
from .word_formation import words_formation_mmocr_tesseract as word_formation
DEFAULT_SETTING_PATH = str(Path(__file__).parents[1]) + "/settings.yml"
class OcrEngine:
def __init__(self, settings_file: str = DEFAULT_SETTING_PATH, **kwargs):
"""Warper of text detection and text recognition
:param settings_file: path to default setting file
:param kwargs: keyword arguments to overwrite the default settings file
"""
with open(settings_file) as f:
# use safe_load instead load
self._settings = yaml.safe_load(f)
self._update_configs(kwargs)
self._ensure_device()
self._detector = StandaloneYOLOXRunner(**self._settings["detector"])
self._recognizer = StandaloneSATRNRunner(**self._settings["recognizer"])
self._deskewer = self._load_deskewer()
def _update_configs(self, params):
for key, para in params.items(): # overwrite default settings by keyword arguments
if key not in self._settings:
raise ValueError("Invalid setting found in OcrEngine: ", k)
if key == "device":
self._settings[key] = para
self._settings["detector"][key] = para
self._settings["recognizer"][key] = para
self._settings["deskew"][key] = para
else:
for k, v in para.items():
if isinstance(v, dict):
for sub_key, sub_value in v.items():
self._settings[key][k][sub_key] = sub_value
else:
self._settings[key][k] = v
def _load_deskewer(self) -> Optional[AlignImage]:
if self._settings["deskew"]["enable"]:
deskewer = AlignImage(
**{k: v for k, v in self._settings["deskew"].items() if k != "enable"}
)
print(
"[WARNING]: Deskew is enabled. The bounding boxes prediction may not be aligned with the original image. In case of using these predictions for pseudo-label, turn on save_deskewed option and use the saved deskewed images instead for further proceed."
)
return deskewer
return None
def _ensure_device(self):
if "cuda" in self._settings["device"]:
if not torch.cuda.is_available():
print("[WARNING]: CUDA is not available, running with cpu instead")
self._settings["device"] = "cpu"
@property
def version(self):
return {
"detector": self._settings["detector"],
"recognizer": self._settings["recognizer"],
}
@property
def settings(self):
return self._settings
# @staticmethod
# def xyxyc_to_xyxy_c(xyxyc: np.ndarray) -> Tuple[List[list], list]:
# '''
# convert sdsvtd yoloX detection output to list of bboxes and list of confidences
# @param xyxyc: array of shape (n, 5)
# '''
# xyxy = xyxyc[:, :4].tolist()
# confs = xyxyc[:, 4].tolist()
# return xyxy, confs
# -> Tuple[np.ndarray, List[Box]]:
def preprocess(self, img: np.ndarray) -> tuple[np.ndarray, bool, float]:
img_ = img.copy()
if self._settings["max_img_size"]:
img_ = mmcv.imrescale(
img,
tuple(self._settings["max_img_size"]),
return_scale=False,
interpolation="bilinear",
backend="cv2",
)
is_blank = False
if self._deskewer:
with Timer("deskew"):
img_, is_blank, angle = self._deskewer(img_)
return img, is_blank, angle # replace img_ to img
# for i, bbox in enumerate(bboxes):
# rotated_bbox = rotate_bbox(bbox, angle, img.shape[:2])
# bboxes[i].bbox = rotated_bbox
return img, is_blank, 0
def run_detect(
self, img: np.ndarray, return_raw: bool = False
) -> Tuple[np.ndarray, Union[List[Box], List[list]]]:
"""
run text detection and return list of xyxyc if return_confidence is True, otherwise return a list of xyxy
"""
pred_det = self._detector(img)
if self._settings["detector"]["auto_rotate"]:
img, pred_det = pred_det
pred_det = pred_det[0] # only image at a time
return (
(img, pred_det.tolist())
if return_raw
else (img, [Box(*xyxyc) for xyxyc in pred_det.tolist()])
)
def run_recog(
self, imgs: List[np.ndarray]
) -> Union[List[str], List[Tuple[str, float]]]:
if len(imgs) == 0:
return list()
pred_rec = self._recognizer(imgs)
return [
(post_process_recog(word), conf)
for word, conf in zip(pred_rec[0], pred_rec[1])
]
def read_img(self, img: str) -> np.ndarray:
return ImageReader.read(img)
def get_cropped_imgs(
self, img: np.ndarray, bboxes: Union[List[Box], List[list]]
) -> Tuple[List[np.ndarray], List[bool]]:
"""
img: np image
bboxes: list of xyxy
"""
lcropped_imgs = list()
mask = list()
for bbox in bboxes:
bbox = Box(*bbox) if isinstance(bbox, list) else bbox
bbox = bbox.get_extend_bbox(self._settings["extend_bbox"])
bbox.clamp_by_img_wh(img.shape[1], img.shape[0])
bbox.to_int()
if not bbox.is_valid():
mask.append(False)
continue
cropped_img = bbox.crop_img(img)
lcropped_imgs.append(cropped_img)
mask.append(True)
return lcropped_imgs, mask
def read_page(
self, img: np.ndarray, bboxes: Union[List[Box], List[list]]
) -> Union[List[WordGroup], List[Line]]:
if len(bboxes) == 0: # no bbox found
return list()
with Timer("cropped imgs"):
lcropped_imgs, mask = self.get_cropped_imgs(img, bboxes)
with Timer("recog"):
# batch_mode for efficiency
pred_recs = self.run_recog(lcropped_imgs)
with Timer("construct words"):
lwords = list()
for i in range(len(pred_recs)):
if not mask[i]:
continue
text, conf_rec = pred_recs[i][0], pred_recs[i][1]
bbox = Box(*bboxes[i]) if isinstance(bboxes[i], list) else bboxes[i]
lwords.append(
Word(
image=img,
text=text,
conf_cls=conf_rec,
bbox_obj=bbox,
conf_detect=bbox._conf,
)
)
with Timer("word formation"):
return word_formation(
lwords, img.shape[1], **self._settings["words_to_lines"]
)[0]
# https://stackoverflow.com/questions/48127642/incompatible-types-in-assignment-on-union
@overload
def __call__(self, img: Union[str, np.ndarray, Image.Image]) -> Page:
...
@overload
def __call__(self, img: List[Union[str, np.ndarray, Image.Image]]) -> Document:
...
def __call__(self, img): # type: ignore #ignoring type before implementing batch_mode
"""
Accept an image or list of them, return ocr result as a page or document
"""
with Timer("read image"):
img = ImageReader.read(img)
if self._settings["batch_size"] == 1:
if isinstance(img, list):
if len(img) == 1:
img = img[0] # in case input type is a 1 page pdf
else:
raise AssertionError(
"list input can only be used with batch_mode enabled"
)
img_deskewed, is_blank, angle = self.preprocess(img)
if is_blank:
print(
"[WARNING]: Blank image detected"
) # TODO: should we stop the execution here?
with Timer("detect"):
img_deskewed, bboxes = self.run_detect(img_deskewed)
with Timer("read_page"):
lsegments = self.read_page(img_deskewed, bboxes)
return Page(lsegments, img, img_deskewed if angle != 0 else None)
else:
# lpages = []
# # chunks to reduce memory footprint
# for imgs in chunks(img, self._batch_size):
# # pred_dets = self._detector(imgs)
# # TEMP: use list comprehension because sdsvtd do not support batch mode of text detection
# img = self.preprocess(img)
# img, bboxes = self.run_detect(img)
# for img_, bboxes_ in zip(imgs, bboxes):
# llines = self.read_page(img, bboxes_)
# page = Page(llines, img)
# lpages.append(page)
# return Document(lpages)
raise NotImplementedError("Batch mode is currently not supported")
if __name__ == "__main__":
img_path = "/mnt/ssd1T/hungbnt/Cello/data/PH/Sea7/Sea_7_1.jpg"
engine = OcrEngine(device="cuda:0")
# https://stackoverflow.com/questions/66435480/overload-following-optional-argument
page = engine(img_path) # type: ignore
print(page._word_segments)

View File

@ -1,369 +0,0 @@
from PIL import ImageFont, ImageDraw, Image, ImageOps
# import matplotlib.pyplot as plt
import numpy as np
import cv2
import os
import time
from typing import Generator, Union, List, overload, Tuple, Callable
import glob
import math
from pathlib import Path
from pdf2image import convert_from_path
# from deskew import determine_skew
# from jdeskew.estimator import get_angle
# from jdeskew.utility import rotate as jrotate
def post_process_recog(text: str) -> str:
text = text.replace("", " ")
return text
def find_maximum_without_outliers(lst: list[int], threshold: float = 1.):
'''
To find the maximum number in a list while excluding its outlier values, you can follow these steps:
Determine the range within which you consider values as outliers. This can be based on a specific threshold or a statistical measure such as the interquartile range (IQR).
Iterate through the list and filter out the outlier values based on the defined range. Keep track of the non-outlier values.
Find the maximum value among the non-outlier values.
'''
# Calculate the lower and upper boundaries for outliers
q1 = np.percentile(lst, 25)
q3 = np.percentile(lst, 75)
iqr = q3 - q1
lower_bound = q1 - threshold * iqr
upper_bound = q3 + threshold * iqr
# Filter out outlier values
non_outliers = [x for x in lst if lower_bound <= x <= upper_bound]
# Find the maximum value among non-outliers
max_value = max(non_outliers)
return max_value
class Timer:
def __init__(self, name: str) -> None:
self.name = name
def __enter__(self):
self.start_time = time.perf_counter()
return self
def __exit__(self, func: Callable, *args):
self.end_time = time.perf_counter()
self.elapsed_time = self.end_time - self.start_time
print(f"[INFO]: {self.name} took : {self.elapsed_time:.6f} seconds")
# def rotate(
# image: np.ndarray, angle: float, background: Union[int, Tuple[int, int, int]]
# ) -> np.ndarray:
# old_width, old_height = image.shape[:2]
# angle_radian = math.radians(angle)
# width = abs(np.sin(angle_radian) * old_height) + abs(np.cos(angle_radian) * old_width)
# height = abs(np.sin(angle_radian) * old_width) + abs(np.cos(angle_radian) * old_height)
# image_center = tuple(np.array(image.shape[1::-1]) / 2)
# rot_mat = cv2.getRotationMatrix2D(image_center, angle, 1.0)
# rot_mat[1, 2] += (width - old_width) / 2
# rot_mat[0, 2] += (height - old_height) / 2
# return cv2.warpAffine(image, rot_mat, (int(round(height)), int(round(width))), borderValue=background)
# def rotate_bbox(bbox: list, angle: float) -> list:
# # Compute the center point of the bounding box
# cx = bbox[0] + bbox[2] / 2
# cy = bbox[1] + bbox[3] / 2
# # Define the scale factor for the rotated bounding box
# scale = 1.0 # following the deskew and jdeskew function
# angle_radian = math.radians(angle)
# # Obtain the rotation matrix using cv2.getRotationMatrix2D()
# M = cv2.getRotationMatrix2D((cx, cy), angle_radian, scale)
# # Apply the rotation matrix to the four corners of the bounding box
# corners = np.array([[bbox[0], bbox[1]],
# [bbox[0] + bbox[2], bbox[1]],
# [bbox[0] + bbox[2], bbox[1] + bbox[3]],
# [bbox[0], bbox[1] + bbox[3]]], dtype=np.float32)
# rotated_corners = cv2.transform(np.array([corners]), M)[0]
# # Compute the bounding box of the rotated corners
# x = int(np.min(rotated_corners[:, 0]))
# y = int(np.min(rotated_corners[:, 1]))
# w = int(np.max(rotated_corners[:, 0]) - np.min(rotated_corners[:, 0]))
# h = int(np.max(rotated_corners[:, 1]) - np.min(rotated_corners[:, 1]))
# rotated_bbox = [x, y, w, h]
# return rotated_bbox
# def rotate_bbox(bbox: List[int], angle: float, old_shape: Tuple[int, int]) -> List[int]:
# # https://medium.com/@pokomaru/image-and-bounding-box-rotation-using-opencv-python-2def6c39453
# bbox_ = [bbox[0], bbox[1], bbox[2], bbox[1], bbox[2], bbox[3], bbox[0], bbox[3]]
# h, w = old_shape
# cx, cy = (int(w / 2), int(h / 2))
# bbox_tuple = [
# (bbox_[0], bbox_[1]),
# (bbox_[2], bbox_[3]),
# (bbox_[4], bbox_[5]),
# (bbox_[6], bbox_[7]),
# ] # put x and y coordinates in tuples, we will iterate through the tuples and perform rotation
# rotated_bbox = []
# for i, coord in enumerate(bbox_tuple):
# M = cv2.getRotationMatrix2D((cx, cy), angle, 1.0)
# cos, sin = abs(M[0, 0]), abs(M[0, 1])
# newW = int((h * sin) + (w * cos))
# newH = int((h * cos) + (w * sin))
# M[0, 2] += (newW / 2) - cx
# M[1, 2] += (newH / 2) - cy
# v = [coord[0], coord[1], 1]
# adjusted_coord = np.dot(M, v)
# rotated_bbox.insert(i, (adjusted_coord[0], adjusted_coord[1]))
# result = [int(x) for t in rotated_bbox for x in t]
# return [result[i] for i in [0, 1, 2, -1]] # reformat to xyxy
# def deskew(image: np.ndarray) -> Tuple[np.ndarray, float]:
# grayscale = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
# angle = 0.
# try:
# angle = determine_skew(grayscale)
# except Exception:
# pass
# rotated = rotate(image, angle, (0, 0, 0)) if angle else image
# return rotated, angle
# def jdeskew(image: np.ndarray) -> Tuple[np.ndarray, float]:
# angle = 0.
# try:
# angle = get_angle(image)
# except Exception:
# pass
# # TODO: change resize = True and scale the bounding box
# rotated = jrotate(image, angle, resize=False) if angle else image
# return rotated, angle
# def deskew()
class ImageReader:
"""
accept anything, return numpy array image
"""
supported_ext = [".png", ".jpg", ".jpeg", ".pdf", ".gif"]
@staticmethod
def validate_img_path(img_path: str) -> None:
if not os.path.exists(img_path):
raise FileNotFoundError(img_path)
if os.path.isdir(img_path):
raise IsADirectoryError(img_path)
if not Path(img_path).suffix.lower() in ImageReader.supported_ext:
raise NotImplementedError("Not supported extension at {}".format(img_path))
@overload
@staticmethod
def read(img: Union[str, np.ndarray, Image.Image]) -> np.ndarray: ...
@overload
@staticmethod
def read(img: List[Union[str, np.ndarray, Image.Image]]) -> List[np.ndarray]: ...
@overload
@staticmethod
def read(img: str) -> List[np.ndarray]: ... # for pdf or directory
@staticmethod
def read(img):
if isinstance(img, list):
return ImageReader.from_list(img)
elif isinstance(img, str) and os.path.isdir(img):
return ImageReader.from_dir(img)
elif isinstance(img, str) and img.endswith(".pdf"):
return ImageReader.from_pdf(img)
else:
return ImageReader._read(img)
@staticmethod
def from_dir(dir_path: str) -> List[np.ndarray]:
if os.path.isdir(dir_path):
image_files = glob.glob(os.path.join(dir_path, "*"))
return ImageReader.from_list(image_files)
else:
raise NotADirectoryError(dir_path)
@staticmethod
def from_str(img_path: str) -> np.ndarray:
ImageReader.validate_img_path(img_path)
return ImageReader.from_PIL(Image.open(img_path))
@staticmethod
def from_np(img_array: np.ndarray) -> np.ndarray:
return img_array
@staticmethod
def from_PIL(img_pil: Image.Image, transpose=True) -> np.ndarray:
# if img_pil.is_animated:
# raise NotImplementedError("Only static images are supported, animated image found")
if transpose:
img_pil = ImageOps.exif_transpose(img_pil)
if img_pil.mode != "RGB":
img_pil = img_pil.convert("RGB")
return np.array(img_pil)
@staticmethod
def from_list(img_list: List[Union[str, np.ndarray, Image.Image]]) -> List[np.ndarray]:
limgs = list()
for img_path in img_list:
try:
if isinstance(img_path, str):
ImageReader.validate_img_path(img_path)
limgs.append(ImageReader._read(img_path))
except (FileNotFoundError, NotImplementedError, IsADirectoryError) as e:
print("[ERROR]: ", e)
print("[INFO]: Skipping image {}".format(img_path))
return limgs
@staticmethod
def from_pdf(pdf_path: str, start_page: int = 0, end_page: int = 0) -> List[np.ndarray]:
pdf_file = convert_from_path(pdf_path)
if end_page is not None:
end_page = min(len(pdf_file), end_page + 1)
limgs = [np.array(pdf_page) for pdf_page in pdf_file[start_page:end_page]]
return limgs
@staticmethod
def _read(img: Union[str, np.ndarray, Image.Image]) -> np.ndarray:
if isinstance(img, str):
return ImageReader.from_str(img)
elif isinstance(img, Image.Image):
return ImageReader.from_PIL(img)
elif isinstance(img, np.ndarray):
return ImageReader.from_np(img)
else:
raise ValueError("Invalid img argument type: ", type(img))
def get_name(file_path, ext: bool = True):
file_path_ = os.path.basename(file_path)
return file_path_ if ext else os.path.splitext(file_path_)[0]
def construct_file_path(dir, file_path, ext=''):
'''
args:
dir: /path/to/dir
file_path /example_path/to/file.txt
ext = '.json'
return
/path/to/dir/file.json
'''
return os.path.join(
dir, get_name(file_path,
True)) if ext == '' else os.path.join(
dir, get_name(file_path,
False)) + ext
def chunks(lst: list, n: int) -> Generator:
"""
Yield successive n-sized chunks from lst.
https://stackoverflow.com/questions/312443/how-do-i-split-a-list-into-equally-sized-chunks
"""
for i in range(0, len(lst), n):
yield lst[i:i + n]
def read_ocr_result_from_txt(file_path: str) -> Tuple[list, list]:
'''
return list of bounding boxes, list of words
'''
with open(file_path, 'r') as f:
lines = f.read().splitlines()
boxes, words = [], []
for line in lines:
if line == "":
continue
x1, y1, x2, y2, text = line.split("\t")
x1, y1, x2, y2 = int(x1), int(y1), int(x2), int(y2)
if text and text != " ":
words.append(text)
boxes.append((x1, y1, x2, y2))
return boxes, words
def get_xyxywh_base_on_format(bbox, format):
if format == "xywh":
x1, y1, w, h = bbox[0], bbox[1], bbox[2], bbox[3]
x2, y2 = x1 + w, y1 + h
elif format == "xyxy":
x1, y1, x2, y2 = bbox
w, h = x2 - x1, y2 - y1
else:
raise NotImplementedError("Invalid format {}".format(format))
return (x1, y1, x2, y2, w, h)
def get_dynamic_params_for_bbox_of_label(text, x1, y1, w, h, img_h, img_w, font, font_scale_offset=1):
font_scale_factor = img_h / (img_w + img_h) * font_scale_offset
font_scale = w / (w + h) * font_scale_factor # adjust font scale by width height
thickness = int(font_scale_factor) + 1
(text_width, text_height) = cv2.getTextSize(text, font, fontScale=font_scale, thickness=thickness)[0]
text_offset_x = x1
text_offset_y = y1 - thickness
box_coords = ((text_offset_x, text_offset_y + 1), (text_offset_x + text_width - 2, text_offset_y - text_height - 2))
return (font_scale, thickness, text_height, box_coords)
def visualize_bbox_and_label(
img, bboxes, texts, bbox_color=(200, 180, 60),
text_color=(0, 0, 0),
format="xyxy", is_vnese=False, draw_text=True):
ori_img_type = type(img)
if is_vnese:
img = Image.fromarray(img) if ori_img_type is np.ndarray else img
draw = ImageDraw.Draw(img)
img_w, img_h = img.size
font_pil_str = "fonts/arial.ttf"
font_cv2 = cv2.FONT_HERSHEY_SIMPLEX
else:
img_h, img_w = img.shape[0], img.shape[1]
font_cv2 = cv2.FONT_HERSHEY_SIMPLEX
for i in range(len(bboxes)):
text = texts[i] # text = "{}: {:.0f}%".format(LABELS[classIDs[i]], confidences[i]*100)
x1, y1, x2, y2, w, h = get_xyxywh_base_on_format(bboxes[i], format)
font_scale, thickness, text_height, box_coords = get_dynamic_params_for_bbox_of_label(
text, x1, y1, w, h, img_h, img_w, font=font_cv2)
if is_vnese:
font_pil = ImageFont.truetype(font_pil_str, size=text_height) # type: ignore
fdraw_text = draw.text # type: ignore
fdraw_bbox = draw.rectangle # type: ignore
# Pil use different coordinate => y = y+thickness = y-thickness + 2*thickness
arg_text = ((box_coords[0][0], box_coords[1][1]), text)
kwarg_text = {"font": font_pil, "fill": text_color, "width": thickness}
arg_rec = ((x1, y1, x2, y2),)
kwarg_rec = {"outline": bbox_color, "width": thickness}
arg_rec_text = ((box_coords[0], box_coords[1]),)
kwarg_rec_text = {"fill": bbox_color, "width": thickness}
else:
# cv2.rectangle(img, box_coords[0], box_coords[1], color, cv2.FILLED)
# cv2.putText(img, text, (text_offset_x, text_offset_y), font, fontScale=font_scale, color=(50, 0,0), thickness=thickness)
# cv2.rectangle(img, (x1, y1), (x2, y2), color, thickness)
fdraw_text = cv2.putText
fdraw_bbox = cv2.rectangle
arg_text = (img, text, box_coords[0])
kwarg_text = {"fontFace": font_cv2, "fontScale": font_scale, "color": text_color, "thickness": thickness}
arg_rec = (img, (x1, y1), (x2, y2))
kwarg_rec = {"color": bbox_color, "thickness": thickness}
arg_rec_text = (img, box_coords[0], box_coords[1])
kwarg_rec_text = {"color": bbox_color, "thickness": cv2.FILLED}
# draw a bounding box rectangle and label on the img
fdraw_bbox(*arg_rec, **kwarg_rec) # type: ignore
if draw_text:
fdraw_bbox(*arg_rec_text, **kwarg_rec_text) # type: ignore
fdraw_text(*arg_text, **kwarg_text) # type: ignore # text have to put in front of rec_text
return np.array(img) if ori_img_type is np.ndarray and is_vnese else img

View File

@ -1,903 +0,0 @@
from builtins import dict
from .dto import Word, Line, WordGroup, Box
from .utils import find_maximum_without_outliers
import numpy as np
from typing import Optional, List, Tuple, Union
############################################################################################################################################################################################################################
############################################################################################################################################################################################################################
### WORDS TO LINES ALGORITHMS FROM MMOCR AND TESSERACT ###############################################################################################################################################################################
############################################################################################################################################################################################################################
############################################################################################################################################################################################################################
DEGREE_TO_RADIAN_COEF = np.pi / 180
MAX_INT = int(2e10 + 9)
MIN_INT = -MAX_INT
def is_on_same_line(box_a, box_b, min_y_overlap_ratio=0.8):
"""Check if two boxes are on the same line by their y-axis coordinates.
Two boxes are on the same line if they overlap vertically, and the length
of the overlapping line segment is greater than min_y_overlap_ratio * the
height of either of the boxes.
Args:
box_a (list), box_b (list): Two bounding boxes to be checked
min_y_overlap_ratio (float): The minimum vertical overlapping ratio
allowed for boxes in the same line
Returns:
The bool flag indicating if they are on the same line
"""
a_y_min = np.min(box_a[1::2])
b_y_min = np.min(box_b[1::2])
a_y_max = np.max(box_a[1::2])
b_y_max = np.max(box_b[1::2])
# Make sure that box a is always the box above another
if a_y_min > b_y_min:
a_y_min, b_y_min = b_y_min, a_y_min
a_y_max, b_y_max = b_y_max, a_y_max
if b_y_min <= a_y_max:
if min_y_overlap_ratio is not None:
sorted_y = sorted([b_y_min, b_y_max, a_y_max])
overlap = sorted_y[1] - sorted_y[0]
min_a_overlap = (a_y_max - a_y_min) * min_y_overlap_ratio
min_b_overlap = (b_y_max - b_y_min) * min_y_overlap_ratio
return overlap >= min_a_overlap or \
overlap >= min_b_overlap
else:
return True
return False
def merge_bboxes_to_group(bboxes_group, x_sorted_boxes):
merged_bboxes = []
for box_group in bboxes_group:
merged_box = {}
merged_box['text'] = ' '.join(
[x_sorted_boxes[idx]['text'] for idx in box_group])
x_min, y_min = float('inf'), float('inf')
x_max, y_max = float('-inf'), float('-inf')
for idx in box_group:
x_max = max(np.max(x_sorted_boxes[idx]['box'][::2]), x_max)
x_min = min(np.min(x_sorted_boxes[idx]['box'][::2]), x_min)
y_max = max(np.max(x_sorted_boxes[idx]['box'][1::2]), y_max)
y_min = min(np.min(x_sorted_boxes[idx]['box'][1::2]), y_min)
merged_box['box'] = [
x_min, y_min, x_max, y_min, x_max, y_max, x_min, y_max
]
merged_box['list_words'] = [x_sorted_boxes[idx]['word']
for idx in box_group]
merged_bboxes.append(merged_box)
return merged_bboxes
def stitch_boxes_into_lines(boxes, max_x_dist=10, min_y_overlap_ratio=0.3):
"""Stitch fragmented boxes of words into lines.
Note: part of its logic is inspired by @Johndirr
(https://github.com/faustomorales/keras-ocr/issues/22)
Args:
boxes (list): List of ocr results to be stitched
max_x_dist (int): The maximum horizontal distance between the closest
edges of neighboring boxes in the same line
min_y_overlap_ratio (float): The minimum vertical overlapping ratio
allowed for any pairs of neighboring boxes in the same line
Returns:
merged_boxes(List[dict]): List of merged boxes and texts
"""
if len(boxes) <= 1:
if len(boxes) == 1:
boxes[0]["list_words"] = [boxes[0]["word"]]
return boxes
# merged_groups = []
merged_lines = []
# sort groups based on the x_min coordinate of boxes
x_sorted_boxes = sorted(boxes, key=lambda x: np.min(x['box'][::2]))
# store indexes of boxes which are already parts of other lines
skip_idxs = set()
i = 0
# locate lines of boxes starting from the leftmost one
for i in range(len(x_sorted_boxes)):
if i in skip_idxs:
continue
# the rightmost box in the current line
rightmost_box_idx = i
line = [rightmost_box_idx]
for j in range(i + 1, len(x_sorted_boxes)):
if j in skip_idxs:
continue
if is_on_same_line(x_sorted_boxes[rightmost_box_idx]['box'],
x_sorted_boxes[j]['box'], min_y_overlap_ratio):
line.append(j)
skip_idxs.add(j)
rightmost_box_idx = j
# split line into lines if the distance between two neighboring
# sub-lines' is greater than max_x_dist
# groups = []
# line_idx = 0
# groups.append([line[0]])
# for k in range(1, len(line)):
# curr_box = x_sorted_boxes[line[k]]
# prev_box = x_sorted_boxes[line[k - 1]]
# dist = np.min(curr_box['box'][::2]) - np.max(prev_box['box'][::2])
# if dist > max_x_dist:
# line_idx += 1
# groups.append([])
# groups[line_idx].append(line[k])
# # Get merged boxes
merged_line = merge_bboxes_to_group([line], x_sorted_boxes)
merged_lines.extend(merged_line)
# merged_group = merge_bboxes_to_group(groups,x_sorted_boxes)
# merged_groups.extend(merged_group)
merged_lines = sorted(merged_lines, key=lambda x: np.min(x['box'][1::2]))
# merged_groups = sorted(merged_groups, key=lambda x: np.min(x['box'][1::2]))
return merged_lines # , merged_groups
# REFERENCE
# https://vigneshgig.medium.com/bounding-box-sorting-algorithm-for-text-detection-and-object-detection-from-left-to-right-and-top-cf2c523c8a85
# https://huggingface.co/spaces/tomofi/MMOCR/blame/main/mmocr/utils/box_util.py
def words_to_lines_mmocr(words: List[Word], *args) -> Tuple[List[Line], Optional[int]]:
bboxes = [{"box": [w.bbox[0], w.bbox[1], w.bbox[2], w.bbox[1], w.bbox[2], w.bbox[3], w.bbox[0], w.bbox[3]],
"text":w._text, "word":w} for w in words]
merged_lines = stitch_boxes_into_lines(bboxes)
merged_groups = merged_lines # TODO: fix code to return both word group and line
lwords_groups = [WordGroup(list_words=merged_box["list_words"],
text=merged_box["text"],
boundingbox=[merged_box["box"][i] for i in [0, 1, 2, -1]])
for merged_box in merged_groups]
llines = [Line(text=word_group._text, list_word_groups=[word_group], boundingbox=word_group._bbox_obj)
for word_group in lwords_groups]
return llines, None # same format with the origin words_to_lines
# lines = [Line() for merged]
# def most_overlapping_row(rows, top, bottom, y_shift):
# max_overlap = -1
# max_overlap_idx = -1
# for i, row in enumerate(rows):
# row_top, row_bottom = row
# overlap = min(top + y_shift, row_top) - max(bottom + y_shift, row_bottom)
# if overlap > max_overlap:
# max_overlap = overlap
# max_overlap_idx = i
# return max_overlap_idx
def most_overlapping_row(rows, row_words, bottom, top, y_shift, max_row_size, y_overlap_threshold=0.5):
max_overlap = -1
max_overlap_idx = -1
overlapping_rows = []
for i, row in enumerate(rows):
row_bottom, row_top = row
overlap = min(bottom - y_shift[i], row_bottom) - \
max(top - y_shift[i], row_top)
if overlap > max_overlap:
max_overlap = overlap
max_overlap_idx = i
# if at least overlap 1 pixel and not (overlap too much and overlap too little)
if (row_top <= bottom and row_bottom >= top) and not (bottom - top - max_overlap > max_row_size * y_overlap_threshold) and not (max_overlap < max_row_size * y_overlap_threshold):
overlapping_rows.append(i)
# Merge overlapping rows if necessary
if len(overlapping_rows) > 1:
merge_bottom = max(rows[i][0] for i in overlapping_rows)
merge_top = min(rows[i][1] for i in overlapping_rows)
if merge_bottom - merge_top <= max_row_size:
# Merge rows
merged_row = (merge_bottom, merge_top)
merged_words = []
# Remove other overlapping rows
for row_idx in overlapping_rows[:0:-1]: # [1,2,3] -> 3,2
merged_words.extend(row_words[row_idx])
del rows[row_idx]
del row_words[row_idx]
rows[overlapping_rows[0]] = merged_row
row_words[overlapping_rows[0]].extend(merged_words[::-1])
max_overlap_idx = overlapping_rows[0]
if bottom - top - max_overlap > max_row_size * y_overlap_threshold and max_overlap < max_row_size * y_overlap_threshold:
max_overlap_idx = -1
return max_overlap_idx
def stitch_boxes_into_lines_tesseract(words: list[Word], max_running_y_shift: int,
gradient: float, y_overlap_threshold: float) -> Tuple[list[list[Word]], float]:
sorted_words = sorted(words, key=lambda x: x.bbox[0])
rows = []
row_words = []
max_row_size = find_maximum_without_outliers([word.height for word in sorted_words])
running_y_shift = []
for _i, word in enumerate(sorted_words):
bbox, _text = word.bbox, word._text
_x1, y1, _x2, y2 = bbox
bottom, top = y2, y1
max_row_size = max(max_row_size, bottom - top)
overlap_row_idx = most_overlapping_row(
rows, row_words, bottom, top, running_y_shift, max_row_size, y_overlap_threshold)
if overlap_row_idx == -1: # No overlapping row found
new_row = (bottom, top)
rows.append(new_row)
row_words.append([word])
running_y_shift.append(0)
else: # Overlapping row found
row_bottom, row_top = rows[overlap_row_idx]
new_bottom = max(row_bottom, bottom)
new_top = min(row_top, top)
rows[overlap_row_idx] = (new_bottom, new_top)
row_words[overlap_row_idx].append(word)
new_shift = (top + bottom) / 2 - (row_top + row_bottom) / 2
running_y_shift[overlap_row_idx] = min(
gradient * running_y_shift[overlap_row_idx] + (1 - gradient) * new_shift, max_running_y_shift) # update and clamp
# Sort rows and row_texts based on the top y-coordinate
sorted_rows_data = sorted(zip(rows, row_words), key=lambda x: x[0][1])
_sorted_rows_idx, sorted_row_words = zip(*sorted_rows_data)
# /_|<- the perpendicular line of the horizontal line and the skew line of the page
page_skew_dist = sum(running_y_shift) / len(running_y_shift)
return sorted_row_words, page_skew_dist
def construct_word_groups_tesseract(sorted_row_words: list[list[Word]],
max_x_dist: int, page_skew_dist: float) -> list[list[list[Word]]]:
# approximate page_skew_angle by page_skew_dist
corrected_max_x_dist = max_x_dist * abs(np.cos(page_skew_dist * DEGREE_TO_RADIAN_COEF))
constructed_row_word_groups = []
for row_words in sorted_row_words:
lword_groups = []
line_idx = 0
lword_groups.append([row_words[0]])
for k in range(1, len(row_words)):
curr_box = row_words[k].bbox
prev_box = row_words[k - 1].bbox
dist = curr_box[0] - prev_box[2]
if dist > corrected_max_x_dist:
line_idx += 1
lword_groups.append([])
lword_groups[line_idx].append(row_words[k])
constructed_row_word_groups.append(lword_groups)
return constructed_row_word_groups
def group_bbox_and_text(lwords: Union[list[Word], list[WordGroup]]) -> tuple[Box, tuple[str, float]]:
text = ' '.join([word._text for word in lwords])
x_min, y_min = MAX_INT, MAX_INT
x_max, y_max = MIN_INT, MIN_INT
conf_det = 0
conf_cls = 0
for word in lwords:
x_max = int(max(np.max(word.bbox[::2]), x_max))
x_min = int(min(np.min(word.bbox[::2]), x_min))
y_max = int(max(np.max(word.bbox[1::2]), y_max))
y_min = int(min(np.min(word.bbox[1::2]), y_min))
conf_det += word._conf_det
conf_cls += word._conf_cls
bbox = Box(x_min, y_min, x_max, y_max, conf=conf_det / len(lwords))
return bbox, (text, conf_cls / len(lwords))
def words_to_lines_tesseract(words: List[Word],
page_width: int, max_running_y_shift_degree: int, gradient: float, max_x_dist: int,
y_overlap_threshold: float) -> Tuple[List[Line],
Optional[float]]:
max_running_y_shift = page_width * np.tan(max_running_y_shift_degree * DEGREE_TO_RADIAN_COEF)
sorted_row_words, page_skew_dist = stitch_boxes_into_lines_tesseract(
words, max_running_y_shift, gradient, y_overlap_threshold)
constructed_row_word_groups = construct_word_groups_tesseract(
sorted_row_words, max_x_dist, page_skew_dist)
llines = []
for row in constructed_row_word_groups:
lwords_row = []
lword_groups = []
for word_group in row:
bbox_word_group, text_word_group = group_bbox_and_text(word_group)
lwords_row.extend(word_group)
lword_groups.append(
WordGroup(
list_words=word_group, text=text_word_group[0],
conf_cls=text_word_group[1],
boundingbox=bbox_word_group))
bbox_line, text_line = group_bbox_and_text(lwords_row)
llines.append(
Line(
list_word_groups=lword_groups, text=text_line[0],
boundingbox=bbox_line, conf_cls=text_line[1]))
return llines, page_skew_dist
### WORDS TO WORDGROUPS #########################################################################################################################################################################################################################
def merge_overlapping_word_groups(
rows: list[list[int]],
row_words: list[list[Word]],
overlapping_rows: list[int],
max_row_size: int) -> bool:
# Merge found overlapping rows if necessary
merge_top = max(rows[i][1] for i in overlapping_rows)
merge_bottom = min(rows[i][3] for i in overlapping_rows)
merge_left = min(rows[i][0] for i in overlapping_rows)
merge_right = max(rows[i][2] for i in overlapping_rows)
if merge_top - merge_bottom <= max_row_size:
# Merge rows
merged_row = [merge_left, merge_top, merge_right, merge_bottom]
merged_words = []
# Remove other overlapping rows
for row_idx in overlapping_rows[:0:-1]: # [1,2,3] -> 3,2
merged_words.extend(row_words[row_idx])
del rows[row_idx]
del row_words[row_idx]
rows[overlapping_rows[0]] = merged_row
row_words[overlapping_rows[0]].extend(merged_words[::-1])
return True
return False
def most_overlapping_word_groups(
rows, row_words, curr_word_bbox, y_shift, max_row_size, y_overlap_threshold, max_x_dist):
max_overlap = -1
max_overlap_idx = -1
overlapping_rows = []
left, top, right, bottom = curr_word_bbox
for i, row in enumerate(rows):
row_left, row_top, row_right, row_bottom = row
top_shift = top - y_shift[i]
bottom_shift = bottom - y_shift[i]
# find the most overlapping row
overlap = min(bottom_shift, row_bottom) - max(top_shift, row_top)
if overlap > max_overlap and min(right - row_left, left - row_right) < max_x_dist:
max_overlap = overlap
max_overlap_idx = i
# exclusive process to handle cases where there are multiple satisfying overlapping rows. For example some rows are not initially overlapping but as the appended words constantly get skewer, there is a change that the end of 1 row would reạch the beginning other row
# if (row_top <= bottom and row_bottom >= top) and not (bottom - top - max_overlap > max_row_size * y_overlap_threshold) and not (max_overlap < max_row_size * y_overlap_threshold):
if (row_top <= bottom_shift and row_bottom >= top_shift) \
and min(right - row_left, left - row_right) < max_x_dist \
and not (bottom - top - overlap > max_row_size * y_overlap_threshold) \
and not (overlap < max_row_size * y_overlap_threshold):
# explain:
# (row_top <= bottom_shift and row_bottom >= top_shift) -> overlap at least 1 pixel
# not (bottom - top - overlap > max_row_size * y_overlap_threshold) -> curr_word is not too big too overlap (to exclude figures containing words)
# not (overlap < max_row_size * y_overlap_threshold) -> overlap too little should not be merged
# min(right - row_left, row_right - left) < max_x_dist -> either the curr_word is close enough to left or right of the curr_row
overlapping_rows.append(i)
if len(overlapping_rows) > 1 and merge_overlapping_word_groups(rows, row_words, overlapping_rows, max_row_size):
max_overlap_idx = overlapping_rows[0]
if bottom - top - max_overlap > max_row_size * y_overlap_threshold and max_overlap < max_row_size * y_overlap_threshold:
max_overlap_idx = -1
return max_overlap_idx
def update_overlapping_word_group_bbox(rows: list[list[int]], overlap_row_idx: int, curr_word_bbox: list[int]) -> None:
left, top, right, bottom = curr_word_bbox
row_left, row_top, row_right, row_bottom = rows[overlap_row_idx]
new_bottom = max(row_bottom, bottom)
new_top = min(row_top, top)
new_left = min(row_left, left)
new_right = max(row_right, right)
rows[overlap_row_idx] = [new_left, new_top, new_right, new_bottom]
def update_word_group_running_y_shift(
running_y_shift: list[float],
overlap_row_idx: int, curr_row_bbox: list[int],
curr_word_bbox: list[int],
gradient: float, max_running_y_shift: float) -> None:
_, top, _, bottom = curr_word_bbox
_, row_top, _, row_bottom = curr_row_bbox
new_shift = (top + bottom) / 2 - (row_top + row_bottom) / 2
running_y_shift[overlap_row_idx] = min(
gradient * running_y_shift[overlap_row_idx] + (1 - gradient) * new_shift, max_running_y_shift) # update and clamp
def stitch_boxes_into_word_groups_tesseract(words: list[Word],
max_running_y_shift: int, gradient: float, y_overlap_threshold: float,
max_x_dist: int) -> Tuple[list[WordGroup], float]:
sorted_words = sorted(words, key=lambda x: x.bbox[0])
rows = []
row_words = []
max_row_size = sorted_words[0].height
running_y_shift = []
for word in sorted_words:
bbox: list[int] = word.bbox
max_row_size = max(max_row_size, bbox[3] - bbox[1])
if bbox[-1] < 200 and word.text == "Nguyễn":
print("DEBUGING")
overlap_row_idx = most_overlapping_word_groups(
rows, row_words, bbox, running_y_shift, max_row_size, y_overlap_threshold, max_x_dist)
if overlap_row_idx == -1: # No overlapping row found
rows.append(bbox) # new row
row_words.append([word]) # new row_word
running_y_shift.append(0)
else: # Overlapping row found
# row_bottom, row_top = rows[overlap_row_idx]
update_overlapping_word_group_bbox(rows, overlap_row_idx, bbox)
row_words[overlap_row_idx].append(word) # update row_words
update_word_group_running_y_shift(
running_y_shift, overlap_row_idx, rows[overlap_row_idx],
bbox, gradient, max_running_y_shift)
# Sort rows and row_texts based on the top y-coordinate
sorted_rows_data = sorted(zip(rows, row_words), key=lambda x: x[0][1])
_sorted_rows_idx, sorted_row_words = zip(*sorted_rows_data)
lword_groups = []
for word_group in sorted_row_words:
bbox_word_group, text_word_group = group_bbox_and_text(word_group)
lword_groups.append(
WordGroup(
list_words=word_group, text=text_word_group[0],
conf_cls=text_word_group[1],
boundingbox=bbox_word_group))
# /_|<- the perpendicular line of the horizontal line and the skew line of the page
page_skew_dist = sum(running_y_shift) / len(running_y_shift)
return lword_groups, page_skew_dist
def is_on_same_line_mmocr_tesseract(box_a: list[int], box_b: list[int], min_y_overlap_ratio: float) -> bool:
a_y_min = box_a[1]
b_y_min = box_b[1]
a_y_max = box_a[3]
b_y_max = box_b[3]
# Make sure that box a is always the box above another
if a_y_min > b_y_min:
a_y_min, b_y_min = b_y_min, a_y_min
a_y_max, b_y_max = b_y_max, a_y_max
if b_y_min <= a_y_max:
if min_y_overlap_ratio is not None:
sorted_y = sorted([b_y_min, b_y_max, a_y_max])
overlap = sorted_y[1] - sorted_y[0]
min_a_overlap = (a_y_max - a_y_min) * min_y_overlap_ratio
min_b_overlap = (b_y_max - b_y_min) * min_y_overlap_ratio
return overlap >= min_a_overlap or \
overlap >= min_b_overlap
else:
return True
return False
def stitch_word_groups_into_lines_mmocr_tesseract(
lword_groups: list[WordGroup],
min_y_overlap_ratio: float) -> list[Line]:
merged_lines = []
# sort groups based on the x_min coordinate of boxes
# store indexes of boxes which are already parts of other lines
sorted_word_groups = sorted(lword_groups, key=lambda x: x.bbox[0])
skip_idxs = set()
i = 0
# locate lines of boxes starting from the leftmost one
for i in range(len(sorted_word_groups)):
if i in skip_idxs:
continue
# the rightmost box in the current line
rightmost_box_idx = i
line = [rightmost_box_idx]
for j in range(i + 1, len(sorted_word_groups)):
if j in skip_idxs:
continue
if is_on_same_line_mmocr_tesseract(sorted_word_groups[rightmost_box_idx].bbox,
sorted_word_groups[j].bbox, min_y_overlap_ratio):
line.append(j)
skip_idxs.add(j)
rightmost_box_idx = j
lword_groups_in_line = [sorted_word_groups[k] for k in line]
bbox_line, text_line = group_bbox_and_text(lword_groups_in_line)
merged_lines.append(
Line(
list_word_groups=lword_groups_in_line, text=text_line[0],
conf_cls=text_line[1],
boundingbox=bbox_line))
merged_lines = sorted(merged_lines, key=lambda x: x.bbox[1])
return merged_lines
def words_formation_mmocr_tesseract(words: List[Word], page_width: int, word_formation_mode: str, max_running_y_shift_degree: int, gradient: float,
max_x_dist: int, y_overlap_threshold: float) -> Tuple[Union[List[WordGroup], list[Line]],
Optional[float]]:
if len(words) == 0:
return [], 0
max_running_y_shift = page_width * np.tan(max_running_y_shift_degree * DEGREE_TO_RADIAN_COEF)
lword_groups, page_skew_dist = stitch_boxes_into_word_groups_tesseract(
words, max_running_y_shift, gradient, y_overlap_threshold, max_x_dist)
if word_formation_mode == "word_group":
return lword_groups, page_skew_dist
elif word_formation_mode == "line":
llines = stitch_word_groups_into_lines_mmocr_tesseract(lword_groups, y_overlap_threshold)
return llines, page_skew_dist
else:
raise NotImplementedError("Word formation mode not supported: {}".format(word_formation_mode))
############################################################################################################################################################################################################################
############################################################################################################################################################################################################################
### END WORDS TO LINES ALGORITHMS FROM MMOCR AND TESSERACT ###############################################################################################################################################################################
############################################################################################################################################################################################################################
############################################################################################################################################################################################################################
# MIN_IOU_HEIGHT = 0.7
# MIN_WIDTH_LINE_RATIO = 0.05
# def resize_to_original(
# boundingbox, scale
# ): # resize coordinates to match size of original image
# left, top, right, bottom = boundingbox
# left *= scale[1]
# right *= scale[1]
# top *= scale[0]
# bottom *= scale[0]
# return [left, top, right, bottom]
# def check_iomin(word: Word, word_group: Word_group):
# min_height = min(
# word.boundingbox[3] - word.boundingbox[1],
# word_group.boundingbox[3] - word_group.boundingbox[1],
# )
# intersect = min(word.boundingbox[3], word_group.boundingbox[3]) - max(
# word.boundingbox[1], word_group.boundingbox[1]
# )
# if intersect / min_height > 0.7:
# return True
# return False
# def prepare_line(words):
# lines = []
# visited = [False] * len(words)
# for id_word, word in enumerate(words):
# if word.invalid_size() == 0:
# continue
# new_line = True
# for i in range(len(lines)):
# if (
# lines[i].in_same_line(word) and not visited[id_word]
# ): # check if word is in the same line with lines[i]
# lines[i].merge_word(word)
# new_line = False
# visited[id_word] = True
# if new_line == True:
# new_line = Line()
# new_line.merge_word(word)
# lines.append(new_line)
# # print(len(lines))
# # sort line from top to bottom according top coordinate
# lines.sort(key=lambda x: x.boundingbox[1])
# return lines
# def __create_word_group(word, word_group_id):
# new_word_group_ = Word_group()
# new_word_group_.list_words = list()
# new_word_group_.word_group_id = word_group_id
# new_word_group_.add_word(word)
# return new_word_group_
# def __sort_line(line):
# line.list_word_groups.sort(
# key=lambda x: x.boundingbox[0]
# ) # sort word in lines from left to right
# return line
# def __merge_text_for_line(line):
# line.text = ""
# for word in line.list_word_groups:
# line.text += " " + word.text
# return line
# def __update_list_word_groups(line, word_group_id, word_id, line_width):
# old_list_word_group = line.list_word_groups
# list_word_groups = []
# inital_word_group = __create_word_group(
# old_list_word_group[0], word_group_id)
# old_list_word_group[0].word_id = word_id
# list_word_groups.append(inital_word_group)
# word_group_id += 1
# word_id += 1
# for word in old_list_word_group[1:]:
# check_word_group = True
# word.word_id = word_id
# word_id += 1
# if (
# (not list_word_groups[-1].text.endswith(":"))
# and (
# (word.boundingbox[0] - list_word_groups[-1].boundingbox[2])
# / line_width
# < MIN_WIDTH_LINE_RATIO
# )
# and check_iomin(word, list_word_groups[-1])
# ):
# list_word_groups[-1].add_word(word)
# check_word_group = False
# if check_word_group:
# new_word_group = __create_word_group(word, word_group_id)
# list_word_groups.append(new_word_group)
# word_group_id += 1
# line.list_word_groups = list_word_groups
# return line, word_group_id, word_id
# def construct_word_groups_in_each_line(lines):
# line_id = 0
# word_group_id = 0
# word_id = 0
# for i in range(len(lines)):
# if len(lines[i].list_word_groups) == 0:
# continue
# # left, top ,right, bottom
# line_width = lines[i].boundingbox[2] - \
# lines[i].boundingbox[0] # right - left
# line_width = 1 # TODO: to remove
# lines[i] = __sort_line(lines[i])
# # update text for lines after sorting
# lines[i] = __merge_text_for_line(lines[i])
# lines[i], word_group_id, word_id = __update_list_word_groups(
# lines[i],
# word_group_id,
# word_id,
# line_width)
# lines[i].update_line_id(line_id)
# line_id += 1
# return lines
# def words_to_lines(words, check_special_lines=True): # words is list of Word instance
# # sort word by top
# words.sort(key=lambda x: (x.boundingbox[1], x.boundingbox[0]))
# # words.sort(key=lambda x: (sum(x.bbox)))
# number_of_word = len(words)
# # print(number_of_word)
# # sort list words to list lines, which have not contained word_group yet
# lines = prepare_line(words)
# # construct word_groups in each line
# lines = construct_word_groups_in_each_line(lines)
# return lines, number_of_word
# def near(word_group1: Word_group, word_group2: Word_group):
# min_height = min(
# word_group1.boundingbox[3] - word_group1.boundingbox[1],
# word_group2.boundingbox[3] - word_group2.boundingbox[1],
# )
# overlap = min(word_group1.boundingbox[3], word_group2.boundingbox[3]) - max(
# word_group1.boundingbox[1], word_group2.boundingbox[1]
# )
# if overlap > 0:
# return True
# if abs(overlap / min_height) < 1.5:
# print("near enough", abs(overlap / min_height), overlap, min_height)
# return True
# return False
# def calculate_iou_and_near(wg1: Word_group, wg2: Word_group):
# min_height = min(
# wg1.boundingbox[3] -
# wg1.boundingbox[1], wg2.boundingbox[3] - wg2.boundingbox[1]
# )
# overlap = min(wg1.boundingbox[3], wg2.boundingbox[3]) - max(
# wg1.boundingbox[1], wg2.boundingbox[1]
# )
# iou = overlap / min_height
# distance = min(
# abs(wg1.boundingbox[0] - wg2.boundingbox[2]),
# abs(wg1.boundingbox[2] - wg2.boundingbox[0]),
# )
# if iou > 0.7 and distance < 0.5 * (wg1.boundingboxp[2] - wg1.boundingbox[0]):
# return True
# return False
# def construct_word_groups_to_kie_label(list_word_groups: list):
# kie_dict = dict()
# for wg in list_word_groups:
# if wg.kie_label == "other":
# continue
# if wg.kie_label not in kie_dict:
# kie_dict[wg.kie_label] = [wg]
# else:
# kie_dict[wg.kie_label].append(wg)
# new_dict = dict()
# for key, value in kie_dict.items():
# if len(value) == 1:
# new_dict[key] = value
# continue
# value.sort(key=lambda x: x.boundingbox[1])
# new_dict[key] = value
# return new_dict
# def invoice_construct_word_groups_to_kie_label(list_word_groups: list):
# kie_dict = dict()
# for wg in list_word_groups:
# if wg.kie_label == "other":
# continue
# if wg.kie_label not in kie_dict:
# kie_dict[wg.kie_label] = [wg]
# else:
# kie_dict[wg.kie_label].append(wg)
# return kie_dict
# def postprocess_total_value(kie_dict):
# if "total_in_words_value" not in kie_dict:
# return kie_dict
# for k, value in kie_dict.items():
# if k == "total_in_words_value":
# continue
# l = []
# for v in value:
# if v.boundingbox[3] <= kie_dict["total_in_words_value"][0].boundingbox[3]:
# l.append(v)
# if len(l) != 0:
# kie_dict[k] = l
# return kie_dict
# def postprocess_tax_code_value(kie_dict):
# if "buyer_tax_code_value" in kie_dict or "seller_tax_code_value" not in kie_dict:
# return kie_dict
# kie_dict["buyer_tax_code_value"] = []
# for v in kie_dict["seller_tax_code_value"]:
# if "buyer_name_key" in kie_dict and (
# v.boundingbox[3] > kie_dict["buyer_name_key"][0].boundingbox[3]
# or near(v, kie_dict["buyer_name_key"][0])
# ):
# kie_dict["buyer_tax_code_value"].append(v)
# continue
# if "buyer_name_value" in kie_dict and (
# v.boundingbox[3] > kie_dict["buyer_name_value"][0].boundingbox[3]
# or near(v, kie_dict["buyer_name_value"][0])
# ):
# kie_dict["buyer_tax_code_value"].append(v)
# continue
# if "buyer_address_value" in kie_dict and near(
# kie_dict["buyer_address_value"][0], v
# ):
# kie_dict["buyer_tax_code_value"].append(v)
# return kie_dict
# def postprocess_tax_code_key(kie_dict):
# if "buyer_tax_code_key" in kie_dict or "seller_tax_code_key" not in kie_dict:
# return kie_dict
# kie_dict["buyer_tax_code_key"] = []
# for v in kie_dict["seller_tax_code_key"]:
# if "buyer_name_key" in kie_dict and (
# v.boundingbox[3] > kie_dict["buyer_name_key"][0].boundingbox[3]
# or near(v, kie_dict["buyer_name_key"][0])
# ):
# kie_dict["buyer_tax_code_key"].append(v)
# continue
# if "buyer_name_value" in kie_dict and (
# v.boundingbox[3] > kie_dict["buyer_name_value"][0].boundingbox[3]
# or near(v, kie_dict["buyer_name_value"][0])
# ):
# kie_dict["buyer_tax_code_key"].append(v)
# continue
# if "buyer_address_value" in kie_dict and near(
# kie_dict["buyer_address_value"][0], v
# ):
# kie_dict["buyer_tax_code_key"].append(v)
# return kie_dict
# def invoice_postprocess(kie_dict: dict):
# # all keys or values which are below total_in_words_value will be thrown away
# kie_dict = postprocess_total_value(kie_dict)
# kie_dict = postprocess_tax_code_value(kie_dict)
# kie_dict = postprocess_tax_code_key(kie_dict)
# return kie_dict
# def throw_overlapping_words(list_words):
# new_list = [list_words[0]]
# for word in list_words:
# overlap = False
# area = (word.boundingbox[2] - word.boundingbox[0]) * (
# word.boundingbox[3] - word.boundingbox[1]
# )
# for word2 in new_list:
# area2 = (word2.boundingbox[2] - word2.boundingbox[0]) * (
# word2.boundingbox[3] - word2.boundingbox[1]
# )
# xmin_intersect = max(word.boundingbox[0], word2.boundingbox[0])
# xmax_intersect = min(word.boundingbox[2], word2.boundingbox[2])
# ymin_intersect = max(word.boundingbox[1], word2.boundingbox[1])
# ymax_intersect = min(word.boundingbox[3], word2.boundingbox[3])
# if xmax_intersect < xmin_intersect or ymax_intersect < ymin_intersect:
# continue
# area_intersect = (xmax_intersect - xmin_intersect) * (
# ymax_intersect - ymin_intersect
# )
# if area_intersect / area > 0.7 or area_intersect / area2 > 0.7:
# overlap = True
# if overlap == False:
# new_list.append(word)
# return new_list
# def check_iou(box1: Word, box2: Box, threshold=0.9):
# area1 = (box1.boundingbox[2] - box1.boundingbox[0]) * (
# box1.boundingbox[3] - box1.boundingbox[1]
# )
# area2 = (box2.xmax - box2.xmin) * (box2.ymax - box2.ymin)
# xmin_intersect = max(box1.boundingbox[0], box2.xmin)
# ymin_intersect = max(box1.boundingbox[1], box2.ymin)
# xmax_intersect = min(box1.boundingbox[2], box2.xmax)
# ymax_intersect = min(box1.boundingbox[3], box2.ymax)
# if xmax_intersect < xmin_intersect or ymax_intersect < ymin_intersect:
# area_intersect = 0
# else:
# area_intersect = (xmax_intersect - xmin_intersect) * (
# ymax_intersect - ymin_intersect
# )
# union = area1 + area2 - area_intersect
# iou = area_intersect / union
# if iou > threshold:
# return True
# return False

View File

@ -1,147 +0,0 @@
import os
import glob
import cv2
import json
import argparse
import numpy as np
from tqdm import tqdm
from PIL import Image
from datetime import datetime
from sdsvkvu.sources.kvu import KVUEngine
from sdsvkvu.sources.utils import export_kvu_outputs, export_sbt_outputs, draw_kvu_outputs
from sdsvkvu.utils.utils import create_dir, write_to_json, pdf2img
from sdsvkvu.utils.query.vat import export_kvu_for_VAT_invoice, merged_kvu_for_VAT_invoice_for_multi_pages
from sdsvkvu.utils.query.sbt import export_kvu_for_SDSAP, merged_kvu_for_SDSAP_for_multi_pages
from sdsvkvu.utils.query.vtb import export_kvu_for_vietin, merged_kvu_for_vietin_for_multi_pages
from sdsvkvu.utils.query.all import export_kvu_for_all, merged_kvu_for_all_for_multi_pages
from sdsvkvu.utils.query.manulife import export_kvu_for_manulife, merged_kvu_for_manulife_for_multi_pages
from sdsvkvu.utils.query.sbt_v2 import export_kvu_for_SBT, merged_kvu_for_SBT_for_multi_pages
def get_args():
args = argparse.ArgumentParser(description='Main file')
args.add_argument('--img_dir', type=str, required=True,
help='path to input image/directory file')
args.add_argument('--save_dir', type=str, required=True,
help='path to save directory')
args.add_argument('--doc_type', type=str, default="vat",
help='type of document')
args.add_argument('--export_img', type=bool, default=False,
help='export image of output visualization')
args.add_argument('--kvu_params', type=str, required=False, default="")
return args.parse_args()
def load_engine(kwargs) -> KVUEngine:
print('[INFO] Loading Key-Value Understanding model ...')
if not isinstance(kwargs, dict):
kwargs = json.loads(kwargs) if kwargs else {}
engine = KVUEngine(**kwargs)
print("[INFO] Loaded model")
print("[INFO] KVU engine settings: \n", engine._settings)
return engine
def process_img(img_path: str, save_dir: str, engine: KVUEngine, export_all: bool, option: str) -> dict:
assert (engine._settings.mode == 4 and option == "sbt_v2") \
or (engine._settings.mode != 4 and option != "sbt_v2"), \
"[ERROR] Mode (4) has just supported option \"sbt_v2\""
print("="*5, os.path.basename(img_path))
create_dir(save_dir)
fname, img_ext = os.path.splitext(os.path.basename(img_path))
out_ext = ".json"
image, lbbox, lwords, pr_class_words, pr_relations = engine.predict(img_path)
if len(lbbox) != 1:
raise ValueError(
f"Not support to predict each separated window: {len(lbbox)}"
)
for i in range(len(lbbox)):
if engine._settings.mode in range(4):
raw_outputs = export_kvu_outputs(lwords[i], lbbox[i], pr_class_words[i], pr_relations[i], engine._settings.class_names)
elif engine._settings.mode == 4:
raw_outputs = export_sbt_outputs(lwords[i], lbbox[i], pr_class_words[i], pr_relations[i], engine._settings.class_names)
if export_all:
save_path = os.path.join(save_dir, 'kvu_results')
create_dir(save_path)
write_to_json(os.path.join(save_path, fname + out_ext), raw_outputs)
# image = Image.open(img_path)
image = np.array(image)
image = draw_kvu_outputs(image, lbbox[i], pr_class_words[i], pr_relations[i], class_names=engine._settings.class_names)
cv2.imwrite(os.path.join(save_path, fname + img_ext), image)
if option == "vat":
outputs = export_kvu_for_VAT_invoice(raw_outputs)
elif option == "sbt":
outputs = export_kvu_for_SDSAP(raw_outputs)
elif option == "vtb":
outputs = export_kvu_for_vietin(raw_outputs)
elif option == "manulife":
outputs = export_kvu_for_manulife(raw_outputs)
elif option == "sbt_v2":
outputs = export_kvu_for_SBT(raw_outputs)
else:
outputs = export_kvu_for_all(raw_outputs)
write_to_json(os.path.join(save_dir, fname + out_ext), outputs)
return outputs
def process_pdf(pdf_path: str, save_dir: str, engine: KVUEngine, export_all: bool, option: str, n_pages: int = -1) -> dict:
out_ext = ".json"
fname, pdf_ext = os.path.splitext(os.path.basename(pdf_path))
img_dirname = '_'.join([os.path.basename(os.path.dirname(pdf_path)), fname])
img_save_dir = os.path.join(save_dir, img_dirname)
create_dir(img_save_dir)
list_img_files = pdf2img(pdf_path, img_save_dir, n_pages=n_pages, return_fname=True)
outputs = []
for img_path in list_img_files:
print("=====", os.path.basename(img_path))
_outputs = process_img(img_path, img_save_dir, engine, export_all=export_all, option=option)
outputs.append(_outputs)
if option == "vat":
outputs = merged_kvu_for_VAT_invoice_for_multi_pages(outputs)
elif option == "sbt":
outputs = merged_kvu_for_SDSAP_for_multi_pages(outputs)
elif option == "vtb":
outputs = merged_kvu_for_vietin_for_multi_pages(outputs)
elif option == "manulife":
outputs = merged_kvu_for_manulife_for_multi_pages(outputs)
elif option == "sbt_v2":
outputs = merged_kvu_for_SBT_for_multi_pages(outputs)
else:
outputs = merged_kvu_for_all_for_multi_pages(outputs)
write_to_json(os.path.join(save_dir, fname + out_ext), outputs)
return outputs
def process_dir(dir_path: str, save_dir: str, engine: KVUEngine, export_all: bool, option: str, dir_level: int = 0) -> None:
list_images = []
for ext in ['JPG', 'PNG', 'jpeg', 'jpg', 'png', 'pdf']:
list_images += glob.glob(os.path.join(dir_path, f"{'*/'*dir_level}*.{ext}"))
print('No. images:', len(list_images))
for file_path in tqdm(list_images):
if os.path.splitext(file_path)[1] == ".pdf":
outputs = process_pdf(file_path, save_dir, engine, export_all=export_all, option=option, n_pages=-1)
else:
outputs = process_img(file_path, save_dir, engine, export_all=export_all, option=option)
def Predictor_KVU(img: str, save_dir: str, engine: KVUEngine) -> dict:
curr_datetime = datetime.now().strftime('%Y-%m-%d %H-%M-%S')
image_path = "/home/thucpd/thucpd/PV2-2023/tmp_image/{}.jpg".format(curr_datetime)
cv2.imwrite(image_path, img)
vat_outputs = process_img(image_path, save_dir, engine, export_all=False, option="vat")
return vat_outputs
if __name__ == "__main__":
args = get_args()
engine = load_engine(args.kvu_params)
# vat_outputs = process_img(args.img_dir, args.save_dir, engine, export_all=True, option="vat")
# vat_outputs = process_pdf(args.img_dir, args.save_dir, engine, export_all=True, option="vat")
process_dir(args.img_dir, args.save_dir, engine, export_all=args.export_img, option=args.doc_type)
print('[INFO] Done')

View File

@ -1,45 +0,0 @@
import os
import torch
from sdsvkvu.model.kvu_model import KVUModel
from sdsvkvu.model.combined_model import ComKVUModel
from sdsvkvu.model.document_kvu_model import DocKVUModel
from sdsvkvu.model.sbt_model import SBTModel
def get_model(cfg):
if cfg.mode == 0 or cfg.mode == 1:
model = ComKVUModel(cfg=cfg)
elif cfg.mode == 2:
model = KVUModel(cfg=cfg)
elif cfg.mode == 3:
model = DocKVUModel(cfg=cfg)
elif cfg.mode == 4:
model = SBTModel(cfg=cfg)
else:
raise ValueError(f'[ERROR] Model mode of {cfg.mode} is not supported')
return model
def load_checkpoint(ckpt_path, model, key_include):
assert os.path.exists(ckpt_path) == True, f"Ckpt path at {ckpt_path} not exist!"
state_dict = torch.load(ckpt_path, 'cpu')['state_dict']
for key in list(state_dict.keys()):
if f'.{key_include}.' not in key:
del state_dict[key]
else:
state_dict[key[4:].replace(key_include + '.', "")] = state_dict[key] # remove net.something.
del state_dict[key]
model.load_state_dict(state_dict, strict=True)
print(f"Load checkpoint at {ckpt_path}")
return model
def load_model_weight(net, pretrained_model_file):
pretrained_model_state_dict = torch.load(pretrained_model_file, map_location="cpu")[
"state_dict"
]
new_state_dict = {}
for k, v in pretrained_model_state_dict.items():
new_k = k
if new_k.startswith("net."):
new_k = new_k[len("net.") :]
new_state_dict[new_k] = v
net.load_state_dict(new_state_dict)

View File

@ -1,71 +0,0 @@
import os
import torch
from torch import nn
from sdsvkvu.model.kvu_model import KVUModel
# from model import load_checkpoint
class ComKVUModel(KVUModel):
def __init__(self, cfg):
super().__init__(cfg)
self.model_cfg = cfg.model
self.freeze = cfg.train.freeze
self.finetune_only = cfg.train.finetune_only
self._get_backbones(self.model_cfg.backbone)
self._create_head()
# if os.path.exists(self.model_cfg.ckpt_model_file):
# self.backbone_layoutxlm = load_checkpoint(self.model_cfg.ckpt_model_file, self.backbone_layoutxlm, 'backbone_layoutxlm')
# self.itc_layer = load_checkpoint(self.model_cfg.ckpt_model_file, self.itc_layer, 'itc_layer')
# self.stc_layer = load_checkpoint(self.model_cfg.ckpt_model_file, self.stc_layer, 'stc_layer')
# self.relation_layer = load_checkpoint(self.model_cfg.ckpt_model_file, self.relation_layer, 'relation_layer')
# self.relation_layer_from_key = load_checkpoint(self.model_cfg.ckpt_model_file, self.relation_layer_from_key, 'relation_layer_from_key')
self.loss_func = nn.CrossEntropyLoss()
# if self.freeze:
# for name, param in self.named_parameters():
# if 'backbone' in name:
# param.requires_grad = False
# if self.finetune_only == 'EE':
# for name, param in self.named_parameters():
# if 'itc_layer' not in name and 'stc_layer' not in name:
# param.requires_grad = False
# if self.finetune_only == 'EL':
# for name, param in self.named_parameters():
# if 'relation_layer' not in name or 'relation_layer_from_key' in name:
# param.requires_grad = False
# if self.finetune_only == 'ELK':
# for name, param in self.named_parameters():
# if 'relation_layer_from_key' not in name:
# param.requires_grad = False
def forward(self, batch):
image = batch["image"]
input_ids_layoutxlm = batch["input_ids_layoutxlm"]
bbox = batch["bbox"]
attention_mask_layoutxlm = batch["attention_mask_layoutxlm"]
backbone_outputs_layoutxlm = self.backbone_layoutxlm(
image=image, input_ids=input_ids_layoutxlm, bbox=bbox, attention_mask=attention_mask_layoutxlm)
last_hidden_states = backbone_outputs_layoutxlm.last_hidden_state[:, :512, :]
last_hidden_states = last_hidden_states.transpose(0, 1).contiguous()
itc_outputs = self.itc_layer(last_hidden_states).transpose(0, 1).contiguous()
stc_outputs = self.stc_layer(last_hidden_states, last_hidden_states).squeeze(0)
el_outputs = self.relation_layer(last_hidden_states, last_hidden_states).squeeze(0)
el_outputs_from_key = self.relation_layer_from_key(last_hidden_states, last_hidden_states).squeeze(0)
head_outputs = {"itc_outputs": itc_outputs, "stc_outputs": stc_outputs,
"el_outputs": el_outputs, "el_outputs_from_key": el_outputs_from_key}
loss = 0.0
if any(['labels' in key for key in batch.keys()]):
loss = self._get_loss(head_outputs, batch)
return head_outputs, loss

View File

@ -1,162 +0,0 @@
import torch
from torch import nn
from sdsvkvu.model.relation_extractor import RelationExtractor
from sdsvkvu.model.kvu_model import KVUModel
# from model import load_checkpoint
class DocKVUModel(KVUModel):
def __init__(self, cfg):
super().__init__(cfg)
self.model_cfg = cfg.model
self.freeze = cfg.train.freeze
self.train_cfg = cfg.train
self.n_classes = len(self.model_cfg.class_names)
self._get_backbones(self.model_cfg.backbone)
self._create_head()
self.loss_func = nn.CrossEntropyLoss()
def _create_head(self):
self.backbone_hidden_size = self.backbone_config.hidden_size
self.head_hidden_size = self.model_cfg.head_hidden_size
self.head_p_dropout = self.model_cfg.head_p_dropout
# self.n_classes = self.model_cfg.n_classes + 1
self.repr_hiddent_size = self.backbone_hidden_size
# (1) Initial token classification
self.itc_layer = nn.Sequential(
nn.Dropout(self.head_p_dropout),
nn.Linear(self.backbone_hidden_size, self.backbone_hidden_size),
nn.Dropout(self.head_p_dropout),
nn.Linear(self.backbone_hidden_size, self.n_classes),
)
# (2) Subsequent token classification
self.stc_layer = RelationExtractor(
n_relations=1, #1
backbone_hidden_size=self.backbone_hidden_size,
head_hidden_size=self.head_hidden_size,
head_p_dropout=self.head_p_dropout,
)
# (3) Linking token classification
self.relation_layer = RelationExtractor(
n_relations=1, #1
backbone_hidden_size=self.backbone_hidden_size,
head_hidden_size=self.head_hidden_size,
head_p_dropout=self.head_p_dropout,
)
# (4) Linking token classification
self.relation_layer_from_key = RelationExtractor(
n_relations=1, #1
backbone_hidden_size=self.backbone_hidden_size,
head_hidden_size=self.head_hidden_size,
head_p_dropout=self.head_p_dropout,
)
# Classfication Layer for whole document
# (1) Initial token classification
self.itc_layer_document = nn.Sequential(
nn.Dropout(self.head_p_dropout),
nn.Linear(self.repr_hiddent_size, self.repr_hiddent_size),
nn.Dropout(self.head_p_dropout),
nn.Linear(self.repr_hiddent_size, self.n_classes),
)
# (2) Subsequent token classification
self.stc_layer_document = RelationExtractor(
n_relations=1,
backbone_hidden_size=self.repr_hiddent_size,
head_hidden_size=self.head_hidden_size,
head_p_dropout=self.head_p_dropout,
)
# (3) Linking token classification
self.relation_layer_document = RelationExtractor(
n_relations=1,
backbone_hidden_size=self.repr_hiddent_size,
head_hidden_size=self.head_hidden_size,
head_p_dropout=self.head_p_dropout,
)
# (4) Linking token classification
self.relation_layer_from_key_document = RelationExtractor(
n_relations=1,
backbone_hidden_size=self.repr_hiddent_size,
head_hidden_size=self.head_hidden_size,
head_p_dropout=self.head_p_dropout,
)
self.itc_layer.apply(self._init_weight)
self.stc_layer.apply(self._init_weight)
self.relation_layer.apply(self._init_weight)
self.relation_layer_from_key.apply(self._init_weight)
self.itc_layer_document.apply(self._init_weight)
self.stc_layer_document.apply(self._init_weight)
self.relation_layer_document.apply(self._init_weight)
self.relation_layer_from_key_document.apply(self._init_weight)
def forward(self, batches):
head_outputs_list = []
loss = 0.0
for batch in batches["windows"]:
image = batch["image"]
input_ids = batch["input_ids_layoutxlm"]
bbox = batch["bbox"]
attention_mask = batch["attention_mask_layoutxlm"]
if self.freeze:
for param in self.backbone.parameters():
param.requires_grad = False
if self.model_cfg.backbone == 'layoutxlm':
backbone_outputs = self.backbone(
image=image, input_ids=input_ids, bbox=bbox, attention_mask=attention_mask
)
else:
backbone_outputs = self.backbone(input_ids, attention_mask=attention_mask)
last_hidden_states = backbone_outputs.last_hidden_state[:, :512, :]
last_hidden_states = last_hidden_states.transpose(0, 1).contiguous()
itc_outputs = self.itc_layer(last_hidden_states).transpose(0, 1).contiguous()
stc_outputs = self.stc_layer(last_hidden_states, last_hidden_states).squeeze(0)
el_outputs = self.relation_layer(last_hidden_states, last_hidden_states).squeeze(0)
el_outputs_from_key = self.relation_layer_from_key(last_hidden_states, last_hidden_states).squeeze(0)
window_repr = last_hidden_states.transpose(0, 1).contiguous()
head_outputs = {"window_repr": window_repr,
"itc_outputs": itc_outputs,
"stc_outputs": stc_outputs,
"el_outputs": el_outputs,
"el_outputs_from_key": el_outputs_from_key}
if any(['labels' in key for key in batch.keys()]):
loss += self._get_loss(head_outputs, batch)
head_outputs_list.append(head_outputs)
batch = batches["documents"]
document_repr = torch.cat([w['window_repr'] for w in head_outputs_list], dim=1)
document_repr = document_repr.transpose(0, 1).contiguous()
itc_outputs = self.itc_layer_document(document_repr).transpose(0, 1).contiguous()
stc_outputs = self.stc_layer_document(document_repr, document_repr).squeeze(0)
el_outputs = self.relation_layer_document(document_repr, document_repr).squeeze(0)
el_outputs_from_key = self.relation_layer_from_key_document(document_repr, document_repr).squeeze(0)
head_outputs = {"itc_outputs": itc_outputs,
"stc_outputs": stc_outputs,
"el_outputs": el_outputs,
"el_outputs_from_key": el_outputs_from_key}
if any(['labels' in key for key in batch.keys()]):
loss += self._get_loss(head_outputs, batch)
return head_outputs, loss

View File

@ -1,300 +0,0 @@
import os
import torch
from torch import nn
from pathlib import Path
from transformers import (
LayoutLMConfig,
LayoutLMModel,
LayoutLMTokenizer,
)
from transformers import (
LayoutLMv2Config,
LayoutLMv2Model,
LayoutLMv2FeatureExtractor,
LayoutXLMTokenizer,
)
from transformers import (
XLMRobertaConfig,
AutoTokenizer,
XLMRobertaModel
)
# from model import load_checkpoint
from sdsvkvu.sources.utils import merged_token_embeddings
from sdsvkvu.model.relation_extractor import RelationExtractor
class KVUModel(nn.Module):
def __init__(self, cfg):
super().__init__()
self.model_cfg = cfg.model
self.freeze = cfg.train.freeze
self.finetune_only = cfg.train.finetune_only
self.n_classes = len(self.model_cfg.class_names)
self._get_backbones(self.model_cfg.backbone)
self._create_head()
# if (cfg.stage == 2) and (os.path.exists(self.model_cfg.ckpt_model_file)):
# self.backbone_layoutxlm = load_checkpoint(self.model_cfg.ckpt_model_file, self.backbone_layoutxlm, 'backbone_layoutxlm')
self._create_head()
self.loss_func = nn.CrossEntropyLoss()
if self.freeze:
for name, param in self.named_parameters():
if "backbone" in name:
param.requires_grad = False
def _create_head(self):
self.backbone_hidden_size = 768
self.head_hidden_size = self.model_cfg.head_hidden_size
self.head_p_dropout = self.model_cfg.head_p_dropout
# self.n_classes = self.model_cfg.n_classes + 1
# (1) Initial token classification
self.itc_layer = nn.Sequential(
nn.Dropout(self.head_p_dropout),
nn.Linear(self.backbone_hidden_size, self.backbone_hidden_size),
nn.Dropout(self.head_p_dropout),
nn.Linear(self.backbone_hidden_size, self.n_classes),
)
# (2) Subsequent token classification
self.stc_layer = RelationExtractor(
n_relations=1, # 1
backbone_hidden_size=self.backbone_hidden_size,
head_hidden_size=self.head_hidden_size,
head_p_dropout=self.head_p_dropout,
)
# (3) Linking token classification
self.relation_layer = RelationExtractor(
n_relations=1, # 1
backbone_hidden_size=self.backbone_hidden_size,
head_hidden_size=self.head_hidden_size,
head_p_dropout=self.head_p_dropout,
)
# (4) Linking token classification
self.relation_layer_from_key = RelationExtractor(
n_relations=1, # 1
backbone_hidden_size=self.backbone_hidden_size,
head_hidden_size=self.head_hidden_size,
head_p_dropout=self.head_p_dropout,
)
self.itc_layer.apply(self._init_weight)
self.stc_layer.apply(self._init_weight)
self.relation_layer.apply(self._init_weight)
# def _get_backbones(self, config_type):
# self.tokenizer_layoutxlm = LayoutXLMTokenizer.from_pretrained('microsoft/layoutxlm-base')
# self.feature_extractor = LayoutLMv2FeatureExtractor(apply_ocr=False)
# self.backbone_layoutxlm = LayoutLMv2Model.from_pretrained('microsoft/layoutxlm-base')
def _get_backbones(self, config_type):
configs = {
"layoutlm": {
"config": LayoutLMConfig,
"tokenizer": LayoutLMTokenizer,
"backbone": LayoutLMModel,
"feature_extrator": LayoutLMv2FeatureExtractor,
},
"layoutxlm": {
"config": LayoutLMv2Config,
"tokenizer": LayoutXLMTokenizer,
"backbone": LayoutLMv2Model,
"feature_extrator": LayoutLMv2FeatureExtractor,
},
"xlm-roberta": {
"config": XLMRobertaConfig,
"tokenizer": AutoTokenizer,
"backbone": XLMRobertaModel,
"feature_extrator": LayoutLMv2FeatureExtractor,
},
}
self.backbone_config = configs[config_type]["config"].from_pretrained(
self.model_cfg.pretrained_model_path
)
if config_type != "xlm-roberta":
self.tokenizer = configs[config_type]["tokenizer"].from_pretrained(
self.model_cfg.pretrained_model_path
)
else:
self.tokenizer = configs[config_type]["tokenizer"].from_pretrained(
self.model_cfg.pretrained_model_path, use_fast=False
)
self.feature_extractor = configs[config_type]["feature_extrator"](
apply_ocr=False
)
self.backbone = configs[config_type]["backbone"].from_pretrained(
self.model_cfg.pretrained_model_path
)
@staticmethod
def _init_weight(module):
init_std = 0.02
if isinstance(module, nn.Linear):
nn.init.normal_(module.weight, 0.0, init_std)
if module.bias is not None:
nn.init.constant_(module.bias, 0.0)
elif isinstance(module, nn.LayerNorm):
nn.init.normal_(module.weight, 1.0, init_std)
if module.bias is not None:
nn.init.constant_(module.bias, 0.0)
def forward(self, lbatches):
windows = lbatches["windows"]
token_embeddings_windows = []
lvalids = []
loverlaps = []
for i, batch in enumerate(windows):
batch = {
k: v.cuda() for k, v in batch.items() if k not in ("img_path", "words")
}
image = batch["image"]
input_ids_layoutxlm = batch["input_ids_layoutxlm"]
bbox = batch["bbox"]
attention_mask_layoutxlm = batch["attention_mask_layoutxlm"]
backbone_outputs_layoutxlm = self.backbone_layoutxlm(
image=image,
input_ids=input_ids_layoutxlm,
bbox=bbox,
attention_mask=attention_mask_layoutxlm,
)
last_hidden_states_layoutxlm = backbone_outputs_layoutxlm.last_hidden_state[
:, :512, :
]
lvalids.append(batch["len_valid_tokens"])
loverlaps.append(batch["len_overlap_tokens"])
token_embeddings_windows.append(last_hidden_states_layoutxlm)
token_embeddings = merged_token_embeddings(
token_embeddings_windows, loverlaps, lvalids, average=False
)
token_embeddings = token_embeddings.transpose(0, 1).contiguous().cuda()
itc_outputs = self.itc_layer(token_embeddings).transpose(0, 1).contiguous()
stc_outputs = self.stc_layer(token_embeddings, token_embeddings).squeeze(0)
el_outputs = self.relation_layer(token_embeddings, token_embeddings).squeeze(0)
el_outputs_from_key = self.relation_layer_from_key(
token_embeddings, token_embeddings
).squeeze(0)
head_outputs = {
"itc_outputs": itc_outputs,
"stc_outputs": stc_outputs,
"el_outputs": el_outputs,
"el_outputs_from_key": el_outputs_from_key,
"embedding_tokens": token_embeddings.transpose(0, 1)
.contiguous()
.detach()
.cpu()
.numpy(),
}
loss = 0.0
if any(["labels" in key for key in lbatches.keys()]):
labels = {
k: v.cuda()
for k, v in lbatches["documents"].items()
if k not in ("img_path")
}
loss = self._get_loss(head_outputs, labels)
return head_outputs, loss
def _get_loss(self, head_outputs, batch):
itc_outputs = head_outputs["itc_outputs"]
stc_outputs = head_outputs["stc_outputs"]
el_outputs = head_outputs["el_outputs"]
el_outputs_from_key = head_outputs["el_outputs_from_key"]
itc_loss = self._get_itc_loss(itc_outputs, batch)
stc_loss = self._get_stc_loss(stc_outputs, batch)
el_loss = self._get_el_loss(el_outputs, batch)
el_loss_from_key = self._get_el_loss(el_outputs_from_key, batch, from_key=True)
loss = itc_loss + stc_loss + el_loss + el_loss_from_key
return loss
def _get_itc_loss(self, itc_outputs, batch):
itc_mask = batch["are_box_first_tokens"].view(-1)
itc_logits = itc_outputs.view(-1, self.model_cfg.n_classes + 1)
itc_logits = itc_logits[itc_mask]
itc_labels = batch["itc_labels"].view(-1)
itc_labels = itc_labels[itc_mask]
itc_loss = self.loss_func(itc_logits, itc_labels)
return itc_loss
def _get_stc_loss(self, stc_outputs, batch):
inv_attention_mask = 1 - batch["attention_mask_layoutxlm"]
bsz, max_seq_length = inv_attention_mask.shape
device = inv_attention_mask.device
invalid_token_mask = torch.cat(
[inv_attention_mask, torch.zeros([bsz, 1]).to(device)], axis=1
).bool()
stc_outputs.masked_fill_(invalid_token_mask[:, None, :], -10000.0)
self_token_mask = (
torch.eye(max_seq_length, max_seq_length + 1).to(device).bool()
)
stc_outputs.masked_fill_(self_token_mask[None, :, :], -10000.0)
stc_mask = batch["attention_mask_layoutxlm"].view(-1).bool()
stc_logits = stc_outputs.view(-1, max_seq_length + 1)
stc_logits = stc_logits[stc_mask]
stc_labels = batch["stc_labels"].view(-1)
stc_labels = stc_labels[stc_mask]
stc_loss = self.loss_func(stc_logits, stc_labels)
return stc_loss
def _get_el_loss(self, el_outputs, batch, from_key=False):
bsz, max_seq_length = batch["attention_mask_layoutxlm"].shape
device = batch["attention_mask_layoutxlm"].device
self_token_mask = (
torch.eye(max_seq_length, max_seq_length + 1).to(device).bool()
)
box_first_token_mask = torch.cat(
[
(batch["are_box_first_tokens"] == False),
torch.zeros([bsz, 1], dtype=torch.bool).to(device),
],
axis=1,
)
el_outputs.masked_fill_(box_first_token_mask[:, None, :], -10000.0)
el_outputs.masked_fill_(self_token_mask[None, :, :], -10000.0)
mask = batch["are_box_first_tokens"].view(-1)
logits = el_outputs.view(-1, max_seq_length + 1)
logits = logits[mask]
if from_key:
el_labels = batch["el_labels_from_key"]
else:
el_labels = batch["el_labels"]
labels = el_labels.view(-1)
labels = labels[mask]
loss = self.loss_func(logits, labels)
return loss

View File

@ -1,48 +0,0 @@
import torch
from torch import nn
class RelationExtractor(nn.Module):
def __init__(
self,
n_relations,
backbone_hidden_size,
head_hidden_size,
head_p_dropout=0.1,
):
super().__init__()
self.n_relations = n_relations
self.backbone_hidden_size = backbone_hidden_size
self.head_hidden_size = head_hidden_size
self.head_p_dropout = head_p_dropout
self.drop = nn.Dropout(head_p_dropout)
self.q_net = nn.Linear(
self.backbone_hidden_size, self.n_relations * self.head_hidden_size
)
self.k_net = nn.Linear(
self.backbone_hidden_size, self.n_relations * self.head_hidden_size
)
self.dummy_node = nn.Parameter(torch.Tensor(1, self.backbone_hidden_size))
nn.init.normal_(self.dummy_node)
def forward(self, h_q, h_k):
h_q = self.q_net(self.drop(h_q))
dummy_vec = self.dummy_node.unsqueeze(0).repeat(1, h_k.size(1), 1)
h_k = torch.cat([h_k, dummy_vec], axis=0)
h_k = self.k_net(self.drop(h_k))
head_q = h_q.view(
h_q.size(0), h_q.size(1), self.n_relations, self.head_hidden_size
)
head_k = h_k.view(
h_k.size(0), h_k.size(1), self.n_relations, self.head_hidden_size
)
relation_score = torch.einsum("ibnd,jbnd->nbij", (head_q, head_k))
return relation_score

View File

@ -1,156 +0,0 @@
import torch
from torch import nn
from transformers import LayoutLMConfig, LayoutLMModel, LayoutLMTokenizer, LayoutLMv2FeatureExtractor
from transformers import LayoutLMv2Config, LayoutLMv2Model
from sdsvkvu.model.relation_extractor import RelationExtractor
from sdsvkvu.model.kvu_model import KVUModel
# from utils import load_checkpoint
class SBTModel(KVUModel):
def __init__(self, cfg):
super().__init__(cfg=cfg)
self.model_cfg = cfg.model
self.freeze = cfg.train.freeze
self.train_cfg = cfg.train
self.n_classes = len(self.model_cfg.class_names)
self._get_backbones(self.model_cfg.backbone)
self._create_head()
self.loss_func = nn.CrossEntropyLoss()
def _create_head(self):
self.backbone_hidden_size = self.backbone_config.hidden_size
self.head_hidden_size = self.model_cfg.head_hidden_size
self.head_p_dropout = self.model_cfg.head_p_dropout
# self.n_classes = self.model_cfg.n_classes + 1
# self.relations = self.model_cfg.n_relations
self.repr_hiddent_size = self.backbone_hidden_size
# (1) Initial token classification
self.itc_layer = nn.Sequential(
nn.Dropout(self.head_p_dropout),
nn.Linear(self.backbone_hidden_size, self.backbone_hidden_size),
nn.Dropout(self.head_p_dropout),
nn.Linear(self.backbone_hidden_size, self.n_classes),
)
# (2) Subsequent token classification
self.stc_layer = RelationExtractor(
n_relations=1, #1
backbone_hidden_size=self.backbone_hidden_size,
head_hidden_size=self.head_hidden_size,
head_p_dropout=self.head_p_dropout,
)
# (3) Linking token classification
self.relation_layer = RelationExtractor(
n_relations=1, #1
backbone_hidden_size=self.backbone_hidden_size,
head_hidden_size=self.head_hidden_size,
head_p_dropout=self.head_p_dropout,
)
# Classfication Layer for whole document
self.itc_layer_document = nn.Sequential(
nn.Dropout(self.head_p_dropout),
nn.Linear(self.repr_hiddent_size, self.repr_hiddent_size),
nn.Dropout(self.head_p_dropout),
nn.Linear(self.repr_hiddent_size, self.n_classes),
)
self.stc_layer_document = RelationExtractor(
n_relations=1,
backbone_hidden_size=self.repr_hiddent_size,
head_hidden_size=self.head_hidden_size,
head_p_dropout=self.head_p_dropout,
)
self.relation_layer_document = RelationExtractor(
n_relations=1,
backbone_hidden_size=self.repr_hiddent_size,
head_hidden_size=self.head_hidden_size,
head_p_dropout=self.head_p_dropout,
)
self.itc_layer.apply(self._init_weight)
self.stc_layer.apply(self._init_weight)
self.relation_layer.apply(self._init_weight)
self.itc_layer_document.apply(self._init_weight)
self.stc_layer_document.apply(self._init_weight)
self.relation_layer_document.apply(self._init_weight)
def forward(self, batches):
head_outputs_list = []
loss = 0.
for batch in batches["windows"]:
image = batch["image"]
input_ids = batch["input_ids_layoutxlm"]
bbox = batch["bbox"]
attention_mask = batch["attention_mask_layoutxlm"]
if self.freeze:
for param in self.backbone.parameters():
param.requires_grad = False
if self.model_cfg.backbone == 'layoutxlm':
backbone_outputs = self.backbone(
image=image, input_ids=input_ids, bbox=bbox, attention_mask=attention_mask
)
else:
backbone_outputs = self.backbone(input_ids, attention_mask=attention_mask)
last_hidden_states = backbone_outputs.last_hidden_state[:, :512, :]
last_hidden_states = last_hidden_states.transpose(0, 1).contiguous()
itc_outputs = self.itc_layer(last_hidden_states).transpose(0, 1).contiguous()
stc_outputs = self.stc_layer(last_hidden_states, last_hidden_states).squeeze(0)
el_outputs = self.relation_layer(last_hidden_states, last_hidden_states).squeeze(0)
window_repr = last_hidden_states.transpose(0, 1).contiguous()
head_outputs = {"window_repr": window_repr,
"itc_outputs": itc_outputs,
"stc_outputs": stc_outputs,
"el_outputs": el_outputs,}
if any(['labels' in key for key in batch.keys()]):
loss += self._get_loss(head_outputs, batch)
head_outputs_list.append(head_outputs)
batch = batches["documents"]
document_repr = torch.cat([w['window_repr'] for w in head_outputs_list], dim=1)
document_repr = document_repr.transpose(0, 1).contiguous()
itc_outputs = self.itc_layer_document(document_repr).transpose(0, 1).contiguous()
stc_outputs = self.stc_layer_document(document_repr, document_repr).squeeze(0)
el_outputs = self.relation_layer_document(document_repr, document_repr).squeeze(0)
head_outputs = {"itc_outputs": itc_outputs,
"stc_outputs": stc_outputs,
"el_outputs": el_outputs}
if any(['labels' in key for key in batch.keys()]):
loss += self._get_loss(head_outputs, batch)
return head_outputs, loss
def _get_loss(self, head_outputs, batch):
itc_outputs = head_outputs["itc_outputs"]
stc_outputs = head_outputs["stc_outputs"]
el_outputs = head_outputs["el_outputs"]
itc_loss = self._get_itc_loss(itc_outputs, batch)
stc_loss = self._get_stc_loss(stc_outputs, batch)
el_loss = self._get_el_loss(el_outputs, batch, from_key=False)
loss = itc_loss + stc_loss + el_loss
return loss

View File

@ -1,225 +0,0 @@
import torch
from pathlib import Path
from omegaconf import OmegaConf
import os
from sdsvkvu.sources.utils import parse_initial_words, parse_subsequent_words, parse_relations
from sdsvkvu.model import get_model, load_model_weight
class KVUPredictor:
def __init__(self, configs):
self.mode = configs.mode
self.device = configs.device
self.pretrained_model_path = configs.model.pretrained_model_path
net, cfg = self._load_model(configs.model.config,
configs.model.checkpoint)
self.model = net
self.class_names = cfg.model.class_names
self.max_seq_length = cfg.train.max_seq_length
self.backbone_type = cfg.model.backbone
if self.mode in (3, 4):
self.slice_interval = 0
self.window_size = cfg.train.window_size
self.max_window_count = cfg.train.max_window_count
self.dummy_idx = self.max_seq_length * self.max_window_count
else:
self.slice_interval = cfg.train.slice_interval
self.window_size = cfg.train.max_num_words
self.max_window_count = 1
if self.mode == 2:
self.dummy_idx = 0 # dynamic dummy
else:
self.dummy_idx = self.max_seq_length # 512
def get_process_configs(self):
_settings = {
# "tokenizer_layoutxlm": self.model.tokenizer_layoutxlm,
# "feature_extractor": self.model.feature_extractor,
"class_names": self.class_names,
"backbone_type": self.backbone_type,
"window_size": self.window_size,
"slice_interval": self.slice_interval,
"max_window_count": self.max_window_count,
"max_seq_length": self.max_seq_length,
"device": self.device,
"mode": self.mode
}
feature_extractor = self.model.feature_extractor
if self.mode in (3, 4):
tokenizer_layoutxlm = self.model.tokenizer
else:
tokenizer_layoutxlm = self.model.tokenizer_layoutxlm
return OmegaConf.create(_settings), tokenizer_layoutxlm, feature_extractor
def _load_model(self, cfg_path, ckpt_path):
cfg = OmegaConf.load(cfg_path)
if self.pretrained_model_path is not None and os.path.exists(self.pretrained_model_path):
cfg.model.pretrained_model_path = self.pretrained_model_path
print("[INFO] Load pretrained backbone at:", cfg.model.pretrained_model_path)
cfg.mode = self.mode
net = get_model(cfg)
load_model_weight(net, ckpt_path)
net.to(self.device)
net.eval()
return net, cfg
def predict(self, input_sample):
if self.mode == 0: # Normal
bbox, lwords, pr_class_words, pr_relations = self.com_predict(input_sample)
return [bbox], [lwords], [pr_class_words], [pr_relations]
elif self.mode == 1: # Full - tokens
bbox, lwords, pr_class_words, pr_relations = self.cat_predict(input_sample)
return [bbox], [lwords], [pr_class_words], [pr_relations]
elif self.mode == 2: # Sliding
bbox, lwords, pr_class_words, pr_relations = [], [], [], []
for window in input_sample['windows']:
_bbox, _lwords, _pr_class_words, _pr_relations = self.com_predict(window)
bbox.append(_bbox)
lwords.append(_lwords)
pr_class_words.append(_pr_class_words)
pr_relations.append(_pr_relations)
return bbox, lwords, pr_class_words, pr_relations
elif self.mode == 3: # Document
bbox, lwords, pr_class_words, pr_relations = self.doc_predict(input_sample)
return [bbox], [lwords], [pr_class_words], [pr_relations]
elif self.mode == 4: # SBT
bbox, lwords, pr_class_words, pr_relations = self.sbt_predict(input_sample)
return [bbox], [lwords], [pr_class_words], [pr_relations]
else:
raise ValueError(f"Not supported mode: {self.mode }")
def doc_predict(self, input_sample):
lwords = input_sample['documents']['words']
for idx, window in enumerate(input_sample['windows']):
input_sample['windows'][idx] = {k: v.unsqueeze(0).to(self.device) for k, v in window.items() if k not in ('words', 'n_empty_windows')}
with torch.no_grad():
head_outputs, _ = self.model(input_sample)
input_sample = input_sample['documents']
head_outputs = {k: v.detach().cpu() for k, v in head_outputs.items()}
# input_sample = {k: v.detach().cpu() for k, v in input_sample.items()}
bbox = input_sample['bbox'].squeeze(0)
pr_class_words, pr_relations = self.kvu_parser(input_sample, head_outputs)
return bbox, lwords, pr_class_words, pr_relations
def com_predict(self, input_sample):
lwords = input_sample['words']
input_sample = {k: v.unsqueeze(0) for k, v in input_sample.items() if k not in ('words', 'img_path')}
input_sample = {k: v.to(self.device) for k, v in input_sample.items()}
with torch.no_grad():
head_outputs, _ = self.model(input_sample)
head_outputs = {k: v.detach().cpu() for k, v in head_outputs.items()}
input_sample = {k: v.detach().cpu() for k, v in input_sample.items()}
bbox = input_sample['bbox'].squeeze(0)
pr_class_words, pr_relations = self.kvu_parser(input_sample, head_outputs)
return bbox, lwords, pr_class_words, pr_relations
def cat_predict(self, input_sample):
lwords = input_sample['documents']['words']
inputs = []
for window in input_sample['windows']:
inputs.append({k: v.unsqueeze(0).cuda() for k, v in window.items() if k not in ('words', 'img_path')})
input_sample['windows'] = inputs
with torch.no_grad():
head_outputs, _ = self.model(input_sample)
head_outputs = {k: v.detach().cpu() for k, v in head_outputs.items() if k not in ('embedding_tokens')}
input_sample = {k: v.unsqueeze(0) for k, v in input_sample["documents"].items()}
bbox = input_sample['bbox'].squeeze(0)
self.dummy_idx = bbox.shape[0]
pr_class_words, pr_relations = self.kvu_parser(input_sample, head_outputs)
return bbox, lwords, pr_class_words, pr_relations
def kvu_parser(self, input_sample, head_outputs):
itc_outputs = head_outputs["itc_outputs"]
stc_outputs = head_outputs["stc_outputs"]
el_outputs = head_outputs["el_outputs"]
el_outputs_from_key = head_outputs["el_outputs_from_key"]
pr_itc_label = torch.argmax(itc_outputs, -1).squeeze(0)
pr_stc_label = torch.argmax(stc_outputs, -1).squeeze(0)
pr_el_label = torch.argmax(el_outputs, -1).squeeze(0)
pr_el_from_key = torch.argmax(el_outputs_from_key, -1).squeeze(0)
box_first_token_mask = input_sample['are_box_first_tokens'].squeeze(0)
attention_mask = input_sample['attention_mask_layoutxlm'].squeeze(0)
pr_init_words = parse_initial_words(pr_itc_label, box_first_token_mask, self.class_names)
pr_class_words = parse_subsequent_words(
pr_stc_label, attention_mask, pr_init_words, self.dummy_idx
)
pr_relations_from_header = parse_relations(pr_el_label, box_first_token_mask, self.dummy_idx)
pr_relations_from_key = parse_relations(pr_el_from_key, box_first_token_mask, self.dummy_idx)
pr_relations = pr_relations_from_header | pr_relations_from_key
return pr_class_words, pr_relations
def sbt_predict(self, input_sample):
lwords = input_sample['documents']['words']
for idx, window in enumerate(input_sample['windows']):
input_sample['windows'][idx] = {k: v.unsqueeze(0).to(self.device) for k, v in window.items() if k not in ('words', 'n_empty_windows')}
with torch.no_grad():
head_outputs, _ = self.model(input_sample)
input_sample = input_sample['documents']
head_outputs = {k: v.detach().cpu() for k, v in head_outputs.items()}
# input_sample = {k: v.detach().cpu() for k, v in input_sample.items()}
bbox = input_sample['bbox'].squeeze(0)
pr_class_words, pr_relations = self.sbt_parser(input_sample, head_outputs)
return bbox, lwords, pr_class_words, pr_relations
def sbt_parser(self, input_sample, head_outputs):
itc_outputs = head_outputs["itc_outputs"]
stc_outputs = head_outputs["stc_outputs"]
el_outputs = head_outputs["el_outputs"]
pr_itc_label = torch.argmax(itc_outputs, -1).squeeze(0)
pr_stc_label = torch.argmax(stc_outputs, -1).squeeze(0)
pr_el_label = torch.argmax(el_outputs, -1).squeeze(0)
box_first_token_mask = input_sample['are_box_first_tokens'].squeeze(0)
attention_mask = input_sample['attention_mask_layoutxlm'].squeeze(0)
pr_init_words = parse_initial_words(pr_itc_label, box_first_token_mask, self.class_names)
pr_class_words = parse_subsequent_words(
pr_stc_label, attention_mask, pr_init_words, self.dummy_idx
)
pr_relations = parse_relations(pr_el_label, box_first_token_mask, self.dummy_idx)
return pr_class_words, pr_relations

View File

@ -1,479 +0,0 @@
import torch
import itertools
import numpy as np
from sdsvkvu.sources.utils import sliding_windows
class KVUProcessor:
def __init__(
self,
tokenizer_layoutxlm,
feature_extractor,
backbone_type,
class_names,
slice_interval,
window_size,
max_seq_length,
mode,
**kwargs,
):
self.mode = mode
self.class_names = class_names
self.backbone_type = backbone_type
self.window_size = window_size
self.slice_interval = slice_interval
self.max_seq_length = max_seq_length
self.tokenizer_layoutxlm = tokenizer_layoutxlm
self.feature_extractor = feature_extractor
self.pad_token_id_layoutxlm = tokenizer_layoutxlm.convert_tokens_to_ids(
tokenizer_layoutxlm._pad_token
)
self.cls_token_id_layoutxlm = tokenizer_layoutxlm.convert_tokens_to_ids(
tokenizer_layoutxlm._cls_token
)
self.sep_token_id_layoutxlm = tokenizer_layoutxlm.convert_tokens_to_ids(
tokenizer_layoutxlm._sep_token
)
self.unk_token_id_layoutxlm = tokenizer_layoutxlm.convert_tokens_to_ids(
self.tokenizer_layoutxlm._unk_token
)
self.class_idx_dic = dict(
[(class_name, idx) for idx, class_name in enumerate(self.class_names)]
)
def __call__(self, lbboxes: list, lwords: list, image, width, height) -> dict:
image = torch.from_numpy(
self.feature_extractor(image)["pixel_values"][0].copy()
)
bbox_windows = sliding_windows(lbboxes, self.window_size, self.slice_interval)
word_windows = sliding_windows(lwords, self.window_size, self.slice_interval)
assert len(bbox_windows) == len(word_windows), \
f"Shape of lbboxes and lwords after sliding window is not the same {len(bbox_windows)} # {len(word_windows)}"
if self.mode == 0: # First 512 tokens
output = self.preprocess_window(
bounding_boxes=lbboxes,
words=lwords,
image_features={"image": image, "width": width, "height": height},
max_seq_length=self.max_seq_length,
)
elif self.mode == 1: # Get full tokens
output = {}
windows = []
for i in range(len(bbox_windows)):
windows.append(
self.preprocess_window(
bounding_boxes=bbox_windows[i],
words=word_windows[i],
image_features={"image": image, "width": width, "height": height},
max_seq_length=self.max_seq_length,
)
)
output["windows"] = windows
elif self.mode == 2: # Sliding window
output = {}
windows = []
output["doduments"] = self.preprocess_window(
bounding_boxes=lbboxes,
words=lwords,
image_features={"image": image, "width": width, "height": height},
max_seq_length=2048,
)
for i in range(len(bbox_windows)):
windows.append(
self.preprocess(
bounding_boxes=bbox_windows[i],
words=word_windows[i],
image_features={"image": image, "width": width, "height": height},
max_seq_length=self.max_seq_length,
)
)
output["windows"] = windows
else:
raise ValueError(f"Not supported mode: {self.mode }")
return output
def preprocess_window(self, bounding_boxes, words, image_features, max_seq_length):
list_word_objects = []
for bb, text in zip(bounding_boxes, words):
boundingBox = [
[bb[0], bb[1]],
[bb[2], bb[1]],
[bb[2], bb[3]],
[bb[0], bb[3]],
]
tokens = self.tokenizer_layoutxlm.convert_tokens_to_ids(
self.tokenizer_layoutxlm.tokenize(text)
)
list_word_objects.append(
{"layoutxlm_tokens": tokens, "boundingBox": boundingBox, "text": text}
)
(
bbox,
input_ids,
attention_mask,
are_box_first_tokens,
box_to_token_indices,
box2token_span_map,
lwords,
len_valid_tokens,
len_non_overlap_tokens,
len_list_tokens,
) = self.parser_words(
list_word_objects,
self.max_seq_length,
image_features["width"],
image_features["height"],
)
assert len_list_tokens == len_valid_tokens + 2
len_overlap_tokens = len_valid_tokens - len_non_overlap_tokens
ntokens = max_seq_length if max_seq_length == 512 else len_valid_tokens + 2
input_ids = input_ids[:ntokens]
attention_mask = attention_mask[:ntokens]
bbox = bbox[:ntokens]
are_box_first_tokens = are_box_first_tokens[:ntokens]
input_ids = torch.from_numpy(input_ids)
attention_mask = torch.from_numpy(attention_mask)
bbox = torch.from_numpy(bbox)
are_box_first_tokens = torch.from_numpy(are_box_first_tokens)
len_valid_tokens = torch.tensor(len_valid_tokens)
len_overlap_tokens = torch.tensor(len_overlap_tokens)
return_dict = {
"words": lwords,
"len_overlap_tokens": len_overlap_tokens,
"len_valid_tokens": len_valid_tokens,
"image": image_features["image"],
"input_ids_layoutxlm": input_ids,
"attention_mask_layoutxlm": attention_mask,
"are_box_first_tokens": are_box_first_tokens,
"bbox": bbox,
}
return return_dict
def parser_words(self, words, max_seq_length, width, height):
list_bbs = []
list_words = []
list_tokens = []
cls_bbs = [0.0] * 8
box2token_span_map = []
box_to_token_indices = []
lwords = [""] * max_seq_length
cum_token_idx = 0
len_valid_tokens = 0
len_non_overlap_tokens = 0
input_ids = np.ones(max_seq_length, dtype=int) * self.pad_token_id_layoutxlm
bbox = np.zeros((max_seq_length, 8), dtype=np.float32)
attention_mask = np.zeros(max_seq_length, dtype=int)
are_box_first_tokens = np.zeros(max_seq_length, dtype=np.bool_)
for word_idx, word in enumerate(words):
this_box_token_indices = []
tokens = word["layoutxlm_tokens"]
bb = word["boundingBox"]
text = word["text"]
len_valid_tokens += len(tokens)
if word_idx < self.slice_interval:
len_non_overlap_tokens += len(tokens)
if len(tokens) == 0:
tokens.append(self.unk_token_id)
if len(list_tokens) + len(tokens) > max_seq_length - 2:
break
box2token_span_map.append(
[len(list_tokens) + 1, len(list_tokens) + len(tokens) + 1]
) # including st_idx
list_tokens += tokens
# min, max clipping
for coord_idx in range(4):
bb[coord_idx][0] = max(0.0, min(bb[coord_idx][0], width))
bb[coord_idx][1] = max(0.0, min(bb[coord_idx][1], height))
bb = list(itertools.chain(*bb))
bbs = [bb for _ in range(len(tokens))]
texts = [text for _ in range(len(tokens))]
for _ in tokens:
cum_token_idx += 1
this_box_token_indices.append(cum_token_idx)
list_bbs.extend(bbs)
list_words.extend(texts) ####
box_to_token_indices.append(this_box_token_indices)
sep_bbs = [width, height] * 4
# For [CLS] and [SEP]
list_tokens = (
[self.cls_token_id_layoutxlm]
+ list_tokens[: max_seq_length - 2]
+ [self.sep_token_id_layoutxlm]
)
if len(list_bbs) == 0:
# When len(json_obj["words"]) == 0 (no OCR result)
list_bbs = [cls_bbs] + [sep_bbs]
else: # len(list_bbs) > 0
list_bbs = [cls_bbs] + list_bbs[: max_seq_length - 2] + [sep_bbs]
# list_words = ['CLS'] + list_words[: max_seq_length - 2] + ['SEP'] ###
# if len(list_words) < 510:
# list_words.extend(['</p>' for _ in range(510 - len(list_words))])
list_words = (
[self.tokenizer_layoutxlm._cls_token]
+ list_words[: max_seq_length - 2]
+ [self.tokenizer_layoutxlm._sep_token]
)
len_list_tokens = len(list_tokens)
input_ids[:len_list_tokens] = list_tokens
attention_mask[:len_list_tokens] = 1
bbox[:len_list_tokens, :] = list_bbs
lwords[:len_list_tokens] = list_words
# Normalize bbox -> 0 ~ 1
bbox[:, [0, 2, 4, 6]] = bbox[:, [0, 2, 4, 6]] / width
bbox[:, [1, 3, 5, 7]] = bbox[:, [1, 3, 5, 7]] / height
if self.backbone_type in ("layoutlm", "layoutxlm"):
bbox = bbox[:, [0, 1, 4, 5]]
bbox = bbox * 1000
bbox = bbox.astype(int)
else:
assert False
st_indices = [
indices[0]
for indices in box_to_token_indices
if indices[0] < max_seq_length
]
are_box_first_tokens[st_indices] = True
return (
bbox,
input_ids,
attention_mask,
are_box_first_tokens,
box_to_token_indices,
box2token_span_map,
lwords,
len_valid_tokens,
len_non_overlap_tokens,
len_list_tokens,
)
class DocKVUProcessor(KVUProcessor):
def __init__(
self,
tokenizer_layoutxlm,
feature_extractor,
backbone_type,
class_names,
max_window_count,
slice_interval,
window_size,
max_seq_length,
mode,
**kwargs,
):
super().__init__(
tokenizer_layoutxlm=tokenizer_layoutxlm,
feature_extractor=feature_extractor,
backbone_type=backbone_type,
class_names=class_names,
slice_interval=slice_interval,
window_size=window_size,
max_seq_length=max_seq_length,
mode=mode,
)
self.max_window_count = max_window_count
self.pad_token_id = self.pad_token_id_layoutxlm
self.cls_token_id = self.cls_token_id_layoutxlm
self.sep_token_id = self.sep_token_id_layoutxlm
self.unk_token_id = self.unk_token_id_layoutxlm
self.tokenizer = self.tokenizer_layoutxlm
def __call__(self, lbboxes: list, lwords: list, images, width, height) -> dict:
image_features = torch.from_numpy(
self.feature_extractor(images)["pixel_values"][0].copy()
)
output = self.preprocess_document(
bounding_boxes=lbboxes,
words=lwords,
image_features={"image": image_features, "width": width, "height": height},
max_seq_length=self.max_seq_length,
)
return output
def preprocess_document(self, bounding_boxes, words, image_features, max_seq_length):
n_words = len(words)
output_dicts = {"windows": [], "documents": []}
n_empty_windows = 0
for i in range(self.max_window_count):
input_ids = np.ones(max_seq_length, dtype=int) * self.pad_token_id
bbox = np.zeros((max_seq_length, 8), dtype=np.float32)
attention_mask = np.zeros(max_seq_length, dtype=int)
are_box_first_tokens = np.zeros(max_seq_length, dtype=np.bool_)
if n_words == 0:
n_empty_windows += 1
output_dicts["windows"].append(
{
"image": image_features["image"],
"input_ids_layoutxlm": torch.from_numpy(input_ids),
"bbox": torch.from_numpy(bbox),
"words": [],
"attention_mask_layoutxlm": torch.from_numpy(attention_mask),
"are_box_first_tokens": torch.from_numpy(are_box_first_tokens),
}
)
continue
start_word_idx = i * self.window_size
stop_word_idx = min(n_words, (i + 1) * self.window_size)
if start_word_idx >= stop_word_idx:
n_empty_windows += 1
output_dicts["windows"].append(output_dicts["windows"][-1])
continue
list_word_objects = []
for bb, text in zip(
bounding_boxes[start_word_idx:stop_word_idx],
words[start_word_idx:stop_word_idx],
):
boundingBox = [
[bb[0], bb[1]],
[bb[2], bb[1]],
[bb[2], bb[3]],
[bb[0], bb[3]],
]
tokens = self.tokenizer_layoutxlm.convert_tokens_to_ids(
self.tokenizer_layoutxlm.tokenize(text)
)
list_word_objects.append(
{
"layoutxlm_tokens": tokens,
"boundingBox": boundingBox,
"text": text,
}
)
(
bbox,
input_ids,
attention_mask,
are_box_first_tokens,
box_to_token_indices,
box2token_span_map,
lwords,
len_valid_tokens,
len_non_overlap_tokens,
len_list_layoutxlm_tokens,
) = self.parser_words(
list_word_objects,
self.max_seq_length,
image_features["width"],
image_features["height"],
)
input_ids = torch.from_numpy(input_ids)
bbox = torch.from_numpy(bbox)
attention_mask = torch.from_numpy(attention_mask)
are_box_first_tokens = torch.from_numpy(are_box_first_tokens)
return_dict = {
"bbox": bbox,
"words": lwords,
"image": image_features["image"],
"input_ids_layoutxlm": input_ids,
"attention_mask_layoutxlm": attention_mask,
"are_box_first_tokens": are_box_first_tokens,
}
output_dicts["windows"].append(return_dict)
attention_mask = torch.cat(
[o["attention_mask_layoutxlm"] for o in output_dicts["windows"]]
)
are_box_first_tokens = torch.cat(
[o["are_box_first_tokens"] for o in output_dicts["windows"]]
)
if n_empty_windows > 0:
attention_mask[
self.max_seq_length * (self.max_window_count - n_empty_windows) :
] = torch.from_numpy(
np.zeros(self.max_seq_length * n_empty_windows, dtype=int)
)
are_box_first_tokens[
self.max_seq_length * (self.max_window_count - n_empty_windows) :
] = torch.from_numpy(
np.zeros(self.max_seq_length * n_empty_windows, dtype=np.bool_)
)
bbox = torch.cat([o["bbox"] for o in output_dicts["windows"]])
words = []
for o in output_dicts["windows"]:
words.extend(o["words"])
return_dict = {
"bbox": bbox,
"words": words,
"attention_mask_layoutxlm": attention_mask,
"are_box_first_tokens": are_box_first_tokens,
"n_empty_windows": n_empty_windows,
}
output_dicts["documents"] = return_dict
return output_dicts
class SBTProcessor(DocKVUProcessor):
def __init__(
self,
tokenizer_layoutxlm,
feature_extractor,
backbone_type,
class_names,
max_window_count,
slice_interval,
window_size,
max_seq_length,
mode,
**kwargs,
):
super().__init__(
tokenizer_layoutxlm,
feature_extractor,
backbone_type,
class_names,
max_window_count,
slice_interval,
window_size,
max_seq_length,
mode,
**kwargs,
)
def __call__(self, lbboxes: list, lwords: list, images, width, height) -> dict:
return super().__call__(lbboxes, lwords, images, width, height)

Some files were not shown because too many files have changed in this diff Show More