Running Experiments

The INI-style experiment configuration file format is slightly updated to allow for future multi-task, multi-lingual setups in terms of data description.

Model-agnostic options are defined in [train] section while the options that will be consumed by the model itself are defined in [model].

An arbitrary number of parallel corpora with multiple languages can be defined in [data] section. Note that you need to define at least train_set and val_set datasets in this section for the training and early-stopping to work correctly.

After creating a configuration file for your own dataset that suits your need, you can run the following command to start training:

nmtpy train -C <config file>

It is possible to override any configuration option through the command-line:

nmtpy train -C <config file> train.<opt>:<val> model.<opt>:<val> ...