[Help Lime] Explain classification results with LIME

[databill86]

Describe the bug Actually this is not a bug. I created a ClassificationModel with 4 classes, and tried to explain the results with LIME. I'm having weird results, so I just want to check if my method was correct.

Here is the code :

class Prediction:

    def __init__(self):

        self.model = ClassificationModel('bert', 'outputs/', use_cuda=False, args={}) 
        self.tokenizer = self.model.tokenizer
        self.max_seq_length = 128
        self.device = "cpu"

    def predictor(self, texts):
        for text in texts: 
            preds, raw_outputs = self.model.predict([text])
            results.append(raw_outputs[0])

        ress = [res for res in results]
        results_array = np.array(ress)
        return results_array 

Then with Lime:

from lime.lime_text import LimeTextExplainer
explainer = LimeTextExplainer(class_names=labels)

prediction = Prediction()
text= "Bueno, me gustaría saber qué talla debo coger"
exp = explainer.explain_instance(text, prediction.predictor, labels=(0,1,2,3), num_features=10, num_samples=len(text.split()))
exp.show_in_notebook(text=True)

I did find some solutions on how to use Lime with BERT, but I guess it is not the same thing with simpletransformers. There is an example here : https://github.com/marcotcr/lime/issues/409

Thank you !

[ThilinaRajapakse]

I haven't tried Lime so I'm afraid I can't give you any concrete examples. What do you mean by weird results though?

[databill86]

The results seem to be random, they change every time I rerun the explainer.

[ThilinaRajapakse]

I can't see anything wrong with the code itself. Do the predictions change every time you run it as well?

[iknoorjobs]

Hi @databill86

Did you figure out the problem? Were you able to run LIME with simple transformers? If yes, please let us know.

Thanks

[databill86]

Hi @iknoorjobs,

No still the same issue, the results are still not stable. I'm not sure I'm doing it right : In the call of explainer.explain_instance, i'm passing num_samples=len(text.split()), I'm not sure if this is right. I did because if was the only way to make it work.

[goru001]

@databill86 @iknoorjobs I think you should increase num_samples to something around 100. The way LIME works is it tries to build interpretable models to explain predictions and those interpretable models are built using num_samples samples. If the num_samples is too low, variance is expected in those interpretable models. It should work out fine after this change.

[iknoorjobs]

@databill86 @iknoorjobs I think you should increase num_samples to something around 100. The way LIME works is it tries to build interpretable models to explain predictions and those interpretable models are built using num_samples samples. If the num_samples is too low, variance is expected in those interpretable models. It should work out fine after this change.

@goru001 Thanks for the update.

[ThilinaRajapakse]

Is this solved now?

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.