-
- -
-

Models and recipes

-
-

Note

-

Currently OpenSeq2Seq has model implementations for machine translation and -automatic speech recognition. -All models work both in float32 and mixed precision. -We recommend you use mixed precision training -when training on Volta GPUs.

-
-

To train models you can use the following commands (don’t forget to substitute -valid config_file path there and number of GPUs if using Horovod).

-

With Horovod (highly recommended when using multiple GPUs):

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

Without Horovod:

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

The description of implemented models is available in the next sections:

-
-

Machine translation

-

The table below contains description and results of -machine translation models available in OpenSeq2Seq. -Currently, we have GNMT-based model, Transformer-based models and -ConvS2S-based models.

-

We measure BLEU score on newstest2014.tok.de file using multi-bleu.perl script from Mosses. -For more details about model descriptions and training setup, -have a look at the configuration files.

-