TensorFlow-Seq2Seq
TensorFlow-Seq2Seq copied to clipboard
tommykwh
Implement en-fr translation task by implenting seq2seq, encoder-decoder in RNN layers with Attention mechanism and Beamsearch inference decoder in TensorFlow 1.3
TensorFlow-Seq2Seq
Implement English-to-French Neural Machine Translation task by implenting seq2seq in TensorFlow 1.3.
Datasets
Shared Task: Machine Translation of News, WMT 16
I took the English and French dataset in news-test2013 as training example.
The hyperparameter and other details in codes are not yet tuned.
Requirements
- Python 3.5
- TensorFlow 1.3
- NLTK
Features
- Tokenization
- Padding (EOS in the end of encoder input; GO in the beginning of decoder input and EOS in the end)
- Vocabulary list building
- Autoencoder/encoder-decoder in multi-RNN-layers
- Attention mechanism (Luong's)
- Batch-training process
- Beamsearch/basic inference decoder
Features In Progress
- Bidirectional encoder
- BLEU score metrics
- Bucketing (Optional. Because TensorFlow's dynamic_decode doesn't require fixed length input anymore)
Usage
Import mySeq2Seq.py to have Seq2SeqModel class.
You may refer to train.ipynb for the training process and data flow.
Reference
Massive Exploration of Neural Machine Translation Architectures
TensorFlow / nmt
JayParks / tf-seq2seq
tommykwh