02_shap.interp.ipynb error?

[ncduy0303]

Hi, I tried running this notebook and got an error at this line:

exp = ShapInterpretation(learn)
exp.decision_plot(class_id=0, row_idx=10)

---

TypeError                                 Traceback (most recent call last)
<ipython-input-12-907d69fd87c2> in <module>
----> 1 exp.decision_plot(class_id=0, row_idx=10)

<ipython-input-8-27630e1fb799> in decision_plot(self, class_id, row_idx, **kwargs)
     16     def decision_plot(self, class_id=0, row_idx=-1, **kwargs):
     17         "Visualize model decision using cumulative `SHAP` values."
---> 18         shap_vals, exp_val = _get_values(self, class_id)
     19         n_rows = shap_vals.shape[0]
     20         if row_idx == -1:

<ipython-input-10-91e394550211> in _get_values(interp, class_id)
      5     exp_vals = interp.explainer.expected_value
      6     if interp.is_multi_output:
----> 7         (class_name, class_idx) = _get_class_info(interp, class_id)
      8         print(f"Classification model detected, displaying score for the class {class_name}.")
      9         print("(use `class_id` to specify another class)")

<ipython-input-9-96fbf3aee7cf> in _get_class_info(interp, class_id)
      2 def _get_class_info(interp:ShapInterpretation, class_id):
      3     "Returns class name associated with index, or vice-versa"
----> 4     if isinstance(class_id, int): class_idx, class_name = class_id, interp.class_names[class_id]
      5     else: class_idx, class_name = interp.class_names.o2i[class_id], class_id
      6     return (class_name, class_idx)

TypeError: 'NoneType' object is not subscriptable

[muellerzr]

I can't seem to reproduce this. What environment are you in?

[ncduy0303]

I'm using Jupyter Notebook on Gradient. Here are the versions of the libraries I'm using:

import shap, fastai2, fastcore, fastinference
fastai2.__version__, fastcore.__version__, fastinference.__version__, shap.__version__
('0.0.18', '0.1.18', '0.0.13', '0.35.0')

[muellerzr]

Thanks! I'll try to do that in gradient today and get back with you, for now I can guarantee it works in Colaboratory (this is where I did my testing real quick)

[ncduy0303]

Thank you!

[muellerzr]

I don't have a gradient account so I'm not sure how to test that actually, but running in my local jupyter environment also doesn't show any issues. The versions are all the same.

[ncduy0303]

Hi, I still saw the error on my side. But after I traced the error to ShapInterpretation(), I changed self.class_names = learn.dl.vocab if hasattr(learn.dl, 'vocab') else None into self.class_names = learn.dls.vocab if hasattr(learn.dls, 'vocab') else None everything seems to work?

#export
class ShapInterpretation():
    "Base interpereter to use the `SHAP` interpretation library"
    def __init__(self, learn:TabularLearner, test_data=None, link='identity', l1_reg='auto', n_samples=128, **kwargs):
        "Initialize `ShapInterpretation` with a Learner, test_data, link, `n_samples`, `l1_reg`, and optional **kwargs"
        self.model = learn.model
        self.dls = learn.dls
        self.class_names = learn.dl.vocab if hasattr(learn.dl, 'vocab') else None # only defined for classification problems
        self.train_data = pd.merge(learn.dls.cats, learn.dls.conts, left_index=True, right_index=True)
        self.test_data = _prepare_data(learn, test_data, n_samples)
        pred_func = partial(_predict, learn)
        self.explainer = shap.SamplingExplainer(pred_func, self.train_data, **kwargs)
        self.shap_vals = self.explainer.shap_values(self.test_data, l1_reg=l1_reg)
        self.is_multi_output = isinstance(self.shap_vals, list)

[muellerzr]

That fix looks alright to me, would you like to put a PR in? :)