https://github.com/Megvii-CSG/MegReader

A research project for text detection and recognition using PyTorch 1.2.
https://github.com/Megvii-CSG/MegReader

ctc deep-learning ocr pytorch text-detection text-detection-recognition text-recognition

Last synced: 7 days ago
JSON representation

A research project for text detection and recognition using PyTorch 1.2.

• Host: GitHub
• URL: https://github.com/Megvii-CSG/MegReader
• Owner: Megvii-CSG
• Archived: true
• Created: 2019-09-28T01:09:23.000Z (about 5 years ago)
• Default Branch: master
• Last Pushed: 2019-12-24T03:56:39.000Z (almost 5 years ago)
• Last Synced: 2024-08-01T13:32:10.712Z (3 months ago)
• Topics: ctc, deep-learning, ocr, pytorch, text-detection, text-detection-recognition, text-recognition
• Language: Python
• Homepage:
• Size: 306 KB
• Stars: 348
• Watchers: 27
• Forks: 67
• Open Issues: 11
• Metadata Files:
  • Readme: README.md
  • Contributing: CONTRIBUTING.md

Awesome Lists containing this project

README

        
# MegReader

A project for research in text detection and recognition using PyTorch 1.2.

This project is originated from the research repo, which heavily relies on closed-source libraries, of CSG-Algorithm team of Megvii(https://megvii.com).

We are in ongoing progress to transfer models into this repo gradually, released implementations are listed in [Progress](#progress).

## Highlights

- Implementations of representative text detection and recognition methods.

- An effective framework for conducting experiments: We use yaml files to configure experiments, making it convenient to take experiments.

- Thorough logging features which make it easy to follow and analyze experimental results.

- CPU/GPU compatible for training and inference.

- Distributed training support.

## Install

### Requirements

`pip install -r requirements.txt`

- Python3.7

- PyTorch 1.2 and CUDA 10.0.

- gcc 5.5(Important for compiling)

### Compile cuda ops (If needed)

```

cd PATH_TO_OPS

python setup.py build_ext --inplace

```

ops may be used:

- DeformableConvV2 `assets/ops/dcn`

- CTC2DLoss `ops/ctc_2d`

### Configuration(optional)

Edit configurations in `config.py`.

## Training

See detailed options: `python3 train.py --help`

## Datasets

We provide data loading implementation with annotation packed with json for quick start. 

Also, lmdb format data are now available too.

You can refer the usage in [demo](experiments/recognition/crnn-lmdb.yaml).

Datasets used in our recognition experiments can be downloaded from [onedrive](https://megvii-my.sharepoint.cn/:f:/g/personal/wanzhaoyi_megvii_com/EjkcrpmiW6hJrUKY-0fEBRABvNMtYniUPfWLVptMmy9-6w?e=bJaYFo). The transform [script](scripts/json_to_lmdb.py) are provide to convert json format data to lmdb.

### Non-distributed

`python3 train.py PATH_TO_EXPERIMENT.yaml --validate --visualize --name NAME_OF_EXPERIMENT`

Following we provide some of configurations of the released recognition models:

- CRNN: `experiments/recognition/crnn.yaml`

- 2D CTC: `experiments/recognition/res50-ppm-2d-ctc.yaml`

- Attention Decoder: `experiments/recognition/fpn50-attention-decoder.yaml`

### Distributed(recommended for multi-gpu training)

`python3 -m torch.distributed.launch --nproc_per_node=NUM_GPUS train.py PATH_TO_EXPERIMENT.yaml -d --validate`

## Evaluating

See detailed options: `python3 eval.py --help`.

Keeping ratio tesing is recommended: `python3 eval.py PATH_TO_EXPERIMENT.yaml --resize_mode keep_ratio`

### Model zoo

Trained models are comming soon.

## Progress

### Recognition Methods

- [x] 2D CTC

- [x] CRNN

- [x] Attention Decoder

- [ ] Rectification

### Detection Methods

- [x] Text Snake

- [x] EAST

### End-to-end

- [ ] Mask Text Spotter

## Contributing

[Contributing.md](CONTRIBUTING.md)