Getting started

OpenSeq2Seq has models for machine translation, speech recogtnition, and speech synthesis. You can find detailed tutorials here:

• Machine Translation
  • Toy task - reversing sequences
  • Creating English-to-German translator
• Speech Recognition
  • How to train the model on LibriSpeech dataset
  • How to build your own language model
• Text-To-Speech
  • How to train the model on LJSpeech dataset

All models can be trained in float32 and mixed precision on Volta GPUs.

For multi-GPU and distribuited training we recommended install Horovod . When training with Horovod, you should use the following commands (don’t forget to substitute valid config_file path there and number of GPUs)

mpiexec --allow-run-as-root -np <num_gpus> python run.py --config_file=... --mode=train_eval --use_horovod=True --enable_logs

To train without Horovod:

python run.py --config_file=... --mode=train_eval --enable_logs