neural-combination-of-HCTR

The realization of neural combination of handwritten Chinese text recognition, the model is mainly adapted from fairseq

Requirement

python 3

pytorch 1.0 or higher

Usage

1. Clone the repo.

   git clone  https://github.com/yudmoe/neural-combination-of-HCTR

2. Using ready-made data

Download the prepared data

semantic_data password：esgg

nosemantic_data password：rb62

semantic_data_withLM password：r69c

These folders include over-segmentation,CRNN and attention-based Chinese handwritten string recognition results

3. Run model

When training, make sure that the path of the training and test data sets in the training parameters is valid

   python threeinput_training.py --PATH1 /trainingdata1 --PATH2 /trainingdata2 --PATH3 /trainingdata3 --testpath1 /testdata1 --testpath2 /testdata2 --testpath3 /testdata3

If the training is interrupted, make sure that the parameters encoder_a_path, encoder_b_path, encoder_c_path, decoder_path correspond to their respective encoders when continuing the training