Unable to understand whether my bert model is predicting this way or there is some issue in my captum implementation

[ammaarahmad1999]

test_iter = iter(test_dataloader)

testing_text = testing_text.to_list()

index = -1

batch = next(test_iter) batch = [r.to(device) for r in batch] input_ids = batch[0] attention_mask = batch[1] index += 1

text = testing_text[index]

print(text)

indices = input_ids[0].detach().tolist() all_tokens = tokenizer.convert_ids_to_tokens(indices)

ref_input_ids = input_ids.clone() ref_input_ids[0, 1:-1] = 0

attributions, delta = lig.attribute(inputs=input_ids, baselines=ref_input_ids, additional_forward_args=(attention_mask), n_steps=70, return_convergence_delta=True)

def summarize_attributions(attributions): attributions = attributions.sum(dim=-1).squeeze(0) attributions = attributions / torch.norm(attributions) return attributions

delta

attributions_sum = summarize_attributions(attributions)

output = model(input_ids, attention_mask)

output

storing couple samples in an array for visualization purposes

score_vis = viz.VisualizationDataRecord(attributions_sum, output[0][0], torch.argmax(output[0]), 1, text, attributions_sum.sum(), all_tokens, delta)

print('\033[1m', 'Visualization For Score', '\033[0m') viz.visualize_text([score_vis])

This is a snapshot of my code above to use captum but the strange thing everytime the focus of attributes is on first word. I don't understand why

[ammaarahmad1999]

Any help regarding this will be helpful

[NarineK]

@ammaarahmad1999, did you check if the attribution scores are exactly the same for the first token ? Also try different classification examples. I don't see a reason why it should always be the same token. There might be also something related to the classifier. Also, try a different classifier.