LIME is killed when explaining a text classification (186 classes) by bert classifier

[SuYanqi]

Hi,

LIME is great and easy to use, but I met a problem when I use LIME to explain a text classification (186 classes) by Bert classifier, it can only run a couple of test entries and then be killed.

I have already set the "num_features=10, num_samples=100, top_labels=10". This problem still happens.

Besides, I find it needs lots of memory. Is it any way to save some memory?

Looking forward to your reply. Thanks

[ptschandl]

As far as I can see the classifier_fn gets all num_samples passed, thus in your case 100 samples end up as a single batch. You can manually change your baseline classifier_fn (assuming you have something like the predictor() function as shown here) so it is using those texts with a batch size of your choice instead and return the (manually) concatenated outputs.

[Achinth04]

i tried LIME for layoutLM for 13 classes and have the same issue bert and layoutLM have similar architecture in the sense they use transformers, i cannot get more than a single word explanation, a single explanation takes 10 gb RAM