kolloldas / tf-transliteration

TensorFlow implementation of Google i18n Transliteration Model

tf-transliteration

TensorFlow implementation of the Google Transliteration model in the paper Sequence-to-sequence neural network models for transliteration. Specifically it implements the BiLSTM-CTC model using Epsilon Insertion.

Usage

For the code to run you need Python 3.5+ and Tensorflow 1.5+

Training: The dataset must be provided as tab separated files. You can get an English-to-Hindi transliteration dataset here Train the model for 10,000 steps, evaluating every 1000 steps:

python transliterate.py --data_file=<filename> --train_steps=10000 --eval_steps=100 --min_eval_frequency=1000

During evaluation the CER will be displayed.

Predicting: For predictions the inputs should be provided as a text file with one example per line.

python transliterate.py --decode_input_file=<filename>

The predictions will be written into a corresponding filename.out.ext

tf-transliteration's Issues

Predicting over a single word

Hi,
I'd like to use your code for predicting over a single word as an input (instead of a txt file) and without converting it to a tfrecord file. Is there a way of doing so?

Thank you in advance

Re-predicting an input file with the same name

When I edit sample.txt file and predict again, sample.out.txt is not overwritten. I have to manually delete sample.out.txt and sample.tfrecord.