sfschouten / court-of-xai Goto Github PK

It seems the entmax function can't handle an alpha of <=1, but it still returns gradients that allow it to get there during training.
This causes the loss to become 'nan'.

Use Tqdm to show how far along the interpretation is for each SaliencyInterpreter. So we can have some idea of how much longer it will take.

• Allow for specifying:
  • directory of trained model to experiment on;
  • which interpreters to include and/or which combinations of interpreters to include;
  • wether to run experiments on cuda;
  • batch size to use during experiment;
• Rename Evaluator to AttentionExperiment[er] to avoid confusion with the evaluate command in AllenNLP.
• [Optional] Add attn_experiment command that starts the experiment from the commandline.

We can use Params.from_file and FromParams.from_params like it is done in allennlp/commands/train.py

AllenNLP 2.0+ introduces some breaking changes we will need to address to upgrade.

The LIME code fails when there's a datareader that concatenates the sequences for DistilBERT. We convert instances back to text, but the concatenated sequence gets converted to a single string, but the datareader expects two separate string sequences to make an Instance.

There's a way to fix this that would also make the code nicer. The implementation of LIME we use assumes unprocessed text, which is why we convert our tokenized Instances back to strings. Instead we should replace some of the LIME code so we can do the replacing of some of the tokens with UNKs ourselves, that way the datareader wouldn't be involved anymore.

This would also have as an added benefit that it would eliminate (possible) mistakes occuring in converting back to raw text and then re-tokenizing. I suspect this is introducing at least some weirdness.

Attention was originally proposed in the context of Machine Translation, so it makes since to include it in our list of tasks.

Components:

• Select dataset
• Add Recurrent seq2seq model
• Add Transformer seq2seq model

Our implementation doesn't work on the pair sequence classifier.