huyhoang17 / MASTER-pytorch

Code for the paper "MASTER: Multi-Aspect Non-local Network for Scene Text Recognition" (Pattern Recognition 2021)

Home Page:

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool


PyTorch reimplementation of "MASTER: Multi-Aspect Non-local Network for Scene Text Recognition" (Pattern Recognition 2021). This project is different from our original implementation that builds on the privacy codebase FastOCR of the company. You can also find Tensorflow reimplementation at MASTER-TF repository, and the performance is almost identical. (PS. Logo inspired by the Master Oogway in Kung Fu Panda)


  • 2021/07: MASTER-mmocr, reimplementation of MASTER by mmocr. @Jiaquan Ye
  • 2021/07: TableMASTER-mmocr, 2nd solution of ICDAR 2021 Competition on Scientific Literature Parsing Task B based on MASTER. @Jiaquan Ye
  • 2021/07: Talk can be found at here (Chinese).
  • 2021/05: Savior, which aims to provide a simple, lightweight, fast integrated, pipelined deployment framework for RPA, is now integrated MASTER for captcha recognition. @Tao Luo
  • 2021/04: Slides can be found at here.

Honors based on MASTER



MASTER is a self-attention based scene text recognizer that (1) not only encodes the input-output attention, but also learns self-attention which encodes feature-feature and target-target relationships inside the encoder and decoder and (2) learns a more powerful and robust intermediate representation to spatial distortion and (3) owns a better training and evaluation efficiency. Overall architecture shown follows.


  • python==3.6
  • torchvision==0.6.1
  • pandas==1.0.5
  • torch==1.5.1
  • numpy==1.16.4
  • tqdm==4.47.0
  • Distance==0.1.3
  • Pillow==7.2.0
pip install -r requirements.txt


Prepare Datasets

  1. Prepare the correct format of files as provided in data folder.
  2. Modify train_dataset and val_dataset args in config.json file, including txt_file, img_root, img_w, img_h.
  3. Modify keys.txt in utils/keys.txt file if needed according to the vocabulary of your dataset.
  4. Modify STRING_MAX_LEN in utils/ file if needed according to the text length of your dataset.

Distributed training with config files

Modify the configurations in configs/config.json and files, then run:


The application will be launched via on a 4 GPU node with one process per GPU (recommend).

This is equivalent to

python -m torch.distributed.launch --nnodes=1 --node_rank=0 --nproc_per_node=4 \
--master_addr= --master_port=5555 \ -c configs/config.json -d 1,2,3,4 --local_world_size 4

and is equivalent to specify indices of available GPUs by CUDA_VISIBLE_DEVICES instead of -d args

CUDA_VISIBLE_DEVICES=1,2,3,4 python -m torch.distributed.launch --nnodes=1 --node_rank=0 --nproc_per_node=4 \
--master_addr= --master_port=5555 \ -c configs/config.json --local_world_size 4

Similarly, it can be launched with a single process that spans all 4 GPUs (if node has 4 available GPUs) using (don't recommend):

CUDA_VISIBLE_DEVICES=1,2,3,4 python -m torch.distributed.launch --nnodes=1 --node_rank=0 --nproc_per_node=1 \
--master_addr= --master_port=5555 \ -c configs/config.json --local_world_size 1

Using Multiple Node

You can enable multi-node multi-GPU training by setting nnodes and node_rank args of the commandline line on every node. e.g., 2 nodes 4 gpus run as follows

Node 1, ip:, then run on node 1 as follows

CUDA_VISIBLE_DEVICES=1,2,3,4 python -m torch.distributed.launch --nnodes=2 --node_rank=0 --nproc_per_node=4 \
--master_addr= --master_port=5555 \ -c configs/config.json --local_world_size 4  

Node 2, ip:, then run on node 2 as follows

CUDA_VISIBLE_DEVICES=2,4,6,7 python -m torch.distributed.launch --nnodes=2 --node_rank=1 --nproc_per_node=4 \
--master_addr= --master_port=5555 \ -c configs/config.json --local_world_size 4  

Debug mode on one GPU/CPU training with config files

This option of training mode can debug code without distributed way. -dist must set to false to turn off distributed mode. -d specify which one gpu will be used.

python -c configs/config.json -d 1 -dist false

Resuming from checkpoints

You can resume from a previously saved checkpoint by:

python -m torch.distributed.launch --nnodes=1 --node_rank=0 --nproc_per_node=4 \
--master_addr= --master_port=5555 \ -d 1,2,3,4 --local_world_size 4 --resume path/to/checkpoint

Finetune from checkpoints

You can finetune from a previously saved checkpoint by:

python -m torch.distributed.launch --nnodes=1 --node_rank=0 --nproc_per_node=4 \
--master_addr= --master_port=5555 \ -d 1,2,3,4 --local_world_size 4 --resume path/to/checkpoint --finetune true

Testing from checkpoints

You can predict from a previously saved checkpoint by:

python --checkpoint path/to/checkpoint --img_folder path/to/img_folder \
               --width 160 --height 48 \
               --output_folder path/to/output_folder \
               --gpu 0 --batch_size 64

Note: width and height must be the same as the settings used during training.


Evaluate squence accuracy and edit distance accuracy:

python utils/ --predict-path predict_result.json --label-path label.txt

Note: label.txt: multi-line, every line containing {ImageFile:<ImageFile>, Label:<TextLabel>}



You can specify the name of the training session in config.json files:

"name": "MASTER_Default",
"run_id": "example"

The checkpoints will be saved in save_dir/name/run_id_timestamp/checkpoint_epoch_n, with timestamp in mmdd_HHMMSS format.

A copy of config.json file will be saved in the same folder.

Note: checkpoints contain:

  'arch': arch,
  'epoch': epoch,
  'model_state_dict': self.model.state_dict(),
  'optimizer': self.optimizer.state_dict(),
  'monitor_best': self.monitor_best,
  'config': self.config

Tensorboard Visualization

This project supports Tensorboard visualization by using either torch.utils.tensorboard or TensorboardX.

  1. Install

    If you are using pytorch 1.1 or higher, install tensorboard by 'pip install tensorboard>=1.14.0'.

    Otherwise, you should install tensorboardx. Follow installation guide in TensorboardX.

  2. Run training

    Make sure that tensorboard option in the config file is turned on.

     "tensorboard" : true
  3. Open Tensorboard server

    Type tensorboard --logdir saved/log/ at the project root, then server will open at http://localhost:6006

By default, values of loss will be logged. If you need more visualizations, use add_scalar('tag', data), add_image('tag', image), etc in the trainer._train_epoch method. add_something() methods in this project are basically wrappers for those of tensorboardX.SummaryWriter and torch.utils.tensorboard.SummaryWriter modules.

Note: You don't have to specify current steps, since WriterTensorboard class defined at logger/ will track current steps.


  • Memory-Cache based Inference


If you find MASTER useful please cite our paper:

  title={{MASTER}: Multi-Aspect Non-local Network for Scene Text Recognition},
  author={Ning Lu and Wenwen Yu and Xianbiao Qi and Yihao Chen and Ping Gong and Rong Xiao and Xiang Bai},
  journal={Pattern Recognition},


This project is licensed under the MIT License. See LICENSE for more details.



Code for the paper "MASTER: Multi-Aspect Non-local Network for Scene Text Recognition" (Pattern Recognition 2021)

License:MIT License


Language:Python 99.6%Language:Shell 0.4%