SeqCLR - Sequence-to-Sequence Contrastive Learning for Text Recognition

A framework for sequence-to-sequence contrastive learning (SeqCLR) of visual representations, applied to text recognition.

A PyTorch implementation of the paper by Aberdam et al. [1].

The implementation is partially based on Clova AI's deep text recognition benchmark.

In the kaggle_notebook folder there are some example scripts I used for training:

• supervised_baseline is to train the raw encoder-decoder without encoder pre-training.
• For the pipeline described in the paper:
  1. pre-train the encoder;
  2. train the decoder on top of the the encoder obtained in the previous step.

N.B.: This is a project I made for an university assignment (my first one in machine learning actually) and I cannot guarantee the correctness and is not in any way meant for any serious application as is. Please review the code before using it.

References

[1] Aberdam, A., R. Litman, S. Tsiper, O. Anschel, R. Slossberg, S. Mazor, R. Manmatha, and P. Perona (2020). Sequence-to-Sequence Contrastive Learning for Text Recognition. DOI: 10.48550/ARXIV.2012.10873.