experiment_utils doesn't allow args to be passed to get_features

[chrishokamp]

some feature extractors, such as the word2vec window feature extractor, have additional keyword args in get_features. experiment_utils currently doesn't allow us to access these args from map_feature_extractor, so we're forced to use the default arguments.

code

[varvara-l]

We need to think of a new feature extractors declaration in config. There is a field "args", which are args for feature extractor init, there should be another set of arguments - for the get_features function of the feature extractor.

Now:

• module: marmot.features.alignment_feature_extractor.AlignmentFeatureExtractor args:
  • ''
  • !join ../examples/word_level_quality_estimation/data/en_es/europarl.1000.en
  • !join ../examples/word_level_quality_estimation/data/en_es/europarl.1000.es

Should be:

• module: marmot.features.alignment_feature_extractor.AlignmentFeatureExtractor args:
  • ''
  • !join ../examples/word_level_quality_estimation/data/en_es/europarl.1000.en
  • !join ../examples/word_level_quality_estimation/data/en_es/europarl.1000.es args_get_features:

[varvara-l]

The question is -- are we going to call get_features with different parameters in one run. In other words, do we already know these parameters during the initialisation, or not.

If we do, they should be specified in init, and removed from get_features keyword arguments.

I can't think of situations when we need to call the same feature extractor once with different arguments, so it seems reasonable to get rid of keyword args.

[varvara-l]

Fixed in all existing features. All parameters of new feature extractors should be specified in init.