Fix transformers

[costrau]

Overview

Closes #3217

There were two problems when using a current transformer model. The first was fixed in #3536 and resulted in all SHAP values being zero. The second problem is that the list of transformer models is hard-coded, so each new model would require a new entry. Fortunately, the transformers library provides such a list, so I changed the code to check against their lists.

Minimal working example (random model, random question)

import transformers
import shap

pmodel = transformers.pipeline("text-classification", "microsoft/phi-2", trust_remote_code=True)

data = [
    "Alice: Who is the current president of the United States?\nBob:"
]

pmodel.model.config.task_specific_params = dict({ "text-generation":{
    "max_new_tokens":20}})

pmodel.model.config.is_decoder = True
exp = shap.Explainer(pmodel, pmodel.tokenizer)
shap_values = exp(data, max_evals=15)
shap.plots.text(shap_values)

Checklist

• All pre-commit checks pass.
• Unit tests added (if fixing a bug or adding a new feature)

[costrau]

The failed test looks like a connection problem to me. Can you rerun the pipeline?

[CloseChoice]

The failed test looks like a connection problem to me. Can you rerun the pipeline?

cc @connortann

[CloseChoice]

Thanks for the PR. I have not checked this thoroughly yet but this looks amazing. Will come back later this week with a more in-depth review.

[codecov]

Codecov Report

Attention: Patch coverage is 17.24138% with 24 lines in your changes are missing coverage. Please review.

Project coverage is 61.45%. Comparing base (74db96c) to head (aa87aa8).
Report is 1 commits behind head on master.

Files	Patch %	Lines
shap/models/_topk_lm.py	4.00%	24 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #3578      +/-   ##
==========================================
- Coverage   61.46%   61.45%   -0.01%     
==========================================
  Files          90       90              
  Lines       12746    12748       +2     
==========================================
  Hits         7834     7834              
- Misses       4912     4914       +2     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.