Pytorch implementation of our paper Adapting OCR with limited labels

• This work was tested with PyTorch 1.2.0, CUDA 9.0, python 3.6 and Ubuntu 16.04.
• requirements can be found in the file.
• Also, please do a pip install pytorch-pretrained-bert as one of our kind contributors pointed out :)
• command to create environment from the file is conda create -n pytorch1.4 --file env.txt
• To activate the environment use: source activate pytorch1.4

• Supervised training

python -m train --name exp1 --path path/to/data

• Main arguments

  • --name: creates a directory where checkpoints will be stored
  • --path: path to dataset.
  • --imgdir: dir name of dataset

• Semi-supervised training

python -m train_semi_supervised --name exp1 --path path --source_dir src_dirname --target_dir tgt_dirname --schedule --noise --alpha=1

• Main arguments
  • --name: creates a directory where checkpoints will be stored
  • --path: path to datasets
  • --source_dir: labelled data directory on which ocr was trained
  • --target_dir: unlabeled data directory on which we want to adapt ocr
  • --percent: percentage of unlabeled data to include in self-training
  • --schedule: will include STLR scheduler while training
  • --train_on_pred: will treat top-predictions as targets
  • --noise: will add gaussian noise to images while training
  • --alpha: set to 1 to include the mixup criterion
  • --combine_scoring: will also take into account the scores outputted by a language model

Note: --combine_scoring works only with line images not word images

• Data
  • Use trdg to generate synthetic data. The script for data generation is included scrips/generate_data.sh.
  • Download two different fonts and keep the data pertaining to each font in source and target dirs.
  • Use one of the fonts to train data from scratch in a supervised manner.
  • Then finetune the trained model on target data using semi-supervised learning
  • A sample lexicon is provided in words.txt. Download different lexicon as per need.

• The OCR architecture is a CNN-LSTM model borrowed from here
• The mixup criterion code is borrowed from here
• STLR is borrowed from this paper