Predictions for unseen data not supported (word-level QE)

[fredblain]

Summary

To produce word-level predictions on test data, one currently has to create artificial dummy labels to use along with the data during training. This is inconvenient as it prevents from using a pretrained model to produce labels for unseen data.

Expected behavior

One should be able to reuse a pretrained word-level QE model to predict word-level labels for unseen data, using the 'sampling' mode.

Related issue(s): #3

[fredblain]

Tested with:

• word-level (BiRNN) QE model (aka 'EncWord');
• backend: Tensorflow.

To produce predictions for unseen data:

• use it as test data;
• add it to 'TEXT_FILES' -- (e.g. TEXT_FILES = {'train': 'train.', 'val': 'dev.', 'test': 'test.'})
• add it to 'EVAL_ON_SETS' -- (e.g. EVAL_ON_SETS = ['test']);
• set 'NO_REF' to True;
• set 'MODE' to 'sampling'.

The path to the file which contains the predictions will be given in the log.