LaTeX OCR

A Pytorch implementation of LinXueyuanStdio/LaTeX_OCR_PRO, forked from qs956/Latex_OCR_Pytorch

It uses CNN as the encoder, and RNN as the decoder

Training

• Prepare the dataset files in .npy format, each element of the ndarray is formatted as:

  {
      'ID': 1,
      'label': "x ^ { 2 } - 1 3 x + 3 6 < 0",
      'image': np.ndarray of shape (width, height, RGB)
  }

• Change the dataset_dir to the folder holding the .npy files

• install extra dependencies

  pip install -r requirements.txt

• start training

  python train.py

Warning

The code is meant for research purpose and is far from production ready.

There're several major drawbacks:

• slow training & inference

  1h/epoch on laptop RTX 2060

• bucketing not implementated, which is presented in the original versions