/OpenSeq2Seq

Multi-GPU sequence to sequence learning

Primary LanguagePythonMIT LicenseMIT

OpenSeq2Seq: multi-gpu sequence to sequence learning

This is a research project, not an official NVIDIA product.

Features

  1. Sequence to sequence learning
    1. Different cell types: LSTM, GRU, GLSTM, SLSTM
    2. Encoders: RNN-based, unidirectional, bi-directional, GNMT-like
    3. Attention mechanisms: Bahdanau, Luong, GNMT-like
    4. Beam search for inference
  2. Single box data parallel multi-gpu training
  3. Distributed (data-parallel) multi-node, mult-gpu training using Horovod
  4. LARS norm scaling algorithm

Documentation