jsam/im2latex

Pytorch implemention of Deep CNN Encoder + LSTM Decoder with Attention for Image to Latex

Im2Latex

Deep CNN Encoder + LSTM Decoder with Attention for Image to Latex, the pytorch implemention of the model architecture used by the Seq2Seq for LaTeX generation

Sample results from this implemention

Experimental results on the IM2LATEX-100K test dataset

BLUE-4	Edit Distance	Exact Match
40.80	44.23	0.27

Getting Started

Install dependency:

pip install -r requirement.txt

Download the dataset for training:

cd data
wget http://lstm.seas.harvard.edu/latex/data/im2latex_validate_filter.lst
wget http://lstm.seas.harvard.edu/latex/data/im2latex_train_filter.lst
wget http://lstm.seas.harvard.edu/latex/data/im2latex_test_filter.lst
wget http://lstm.seas.harvard.edu/latex/data/formula_images_processed.tar.gz
wget http://lstm.seas.harvard.edu/latex/data/im2latex_formulas.norm.lst
tar -zxvf formula_images_processed.tar.gz

Preprocess:

python preprocess.py

Build vocab

python build_vocab.py

Train:

 python train.py \
      --data_path=[data dir] \
      --save_dir=[the dir for saving ckpts] \
      --dropout=0.2 --add_position_features \
      --epoches=25 --max_len=150

Evaluate:

python evaluate.py --split=test \
     --model_path=[the path to model] \
     --data_path=[data dir] \
     --batch_size=32 \
     --ref_path=[the file to store reference] \
     --result_path=[the file to store decoding result]

Features

• Schedule Sampling from Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks
• Positional Embedding from Attention Is All You Need
• Batch beam search
• Training from checkpoint
• Improve the code of data loading for cpu/cuda memery efficiency
• Finetune hyper parameters for better performance
• A HTML Page allowing upload picture to decode