Online Handwriting Recognitin Model built on Tensorflow.

General instructions
The dataset(IamOnDb) in npy format after the preprocess step and read step can be found on DGX-2(Path: /workspace/data/src_hit_word_beam_edition). Otherwise download the dataset and run the preprocess.py and read.py in order.


STEPS TO RUN:

1. Decoder folder contains all the tried decoders:
	a. BS --> Beam Search without LM
	b. BS_LM --> Beam Search with LM
	c. WBS --> Word Beam Search
	d. Greedy --> Best path decoder


	Use the corresponding model file with the Predict file to test
	Run predict_*corresponding_decoder*.py
	Make sure the appropriate model, checkpoints are included.
 

2. corpus folder contains the different corpora used for experimentation. Use them in the appropriate Predict.py file

3. checkpoints folder contains the most updated training checkpoints which have given best results overall. Use the exact name of the checkpoints in the predict methods or while restoring the models.

4. The Random_indices folder(test_files.npy) contains the random set that we chose while training. The file itself contains a random permutation of the lines, of which ~75% are used for training and ~25% for testing.

5. Training folder contains the file required to run the code on training. Use the model.py in greedy decoder for speedy results.

6. CTCWordBeamSearch folder contains the package required to run Word Beam Search. This folder won't be required for other decoding methods.

7. BSPLINE_ON_SAMSUNG contains the files which preprocess samsung dataset and upsample the same by using Bspline fit.