OpenSeq2Seq: multi-gpu sequence to sequence learning
This is a research project, not an official NVIDIA product.
- Sequence to sequence learning
- Different cell types: LSTM, GRU, GLSTM, SLSTM
- Encoders: RNN-based, unidirectional, bi-directional, GNMT-like
- Attention mechanisms: Bahdanau, Luong, GNMT-like
- Beam search for inference
- Data parallel multi-gpu training
- LARS norm scaling algorithm
- Python 3.6
- Tensorflow r1.2 or r1.3 (with GPU support)
- NLTK v3.2.3+
Checkout the code and make sure the following test pass: You should see OK after each test
./create_toy_data.sh python -m unittest test/data_layer_tests.py python -m unittest test/model_tests.py
Full documentation is available on the repository wiki.