Update app.py

app.py

@@ -138,10 +138,16 @@ def run_interpretation(raw_original_prompt, raw_interpretation_prompt, max_new_t
     generation_texts = tokenizer.batch_decode(generated)
 
     # try identifying important layers
-    vectors_to_compare = interpreted_vectors # torch.tensor(global_state.sentence_transformer.encode(generation_texts))
+    # vectors_to_compare = interpreted_vectors # torch.tensor(global_state.sentence_transformer.encode(generation_texts))
+    # diff_score = F.normalize(vectors_to_compare, dim=-1).diff(dim=0).norm(dim=-1)
+    bags_of_words = [set(tokenizer.tokenize(text)) for text in generation_texts]
+    diff_score = torch.tensor([
+                                len(bags_of_words[i+1] & bags_of_words[i]) / np.sqrt(len(bags_of_words[i+1]) * len(bags_of_words[i]))
+                                for i in range(len(bags_of_words)-1)
+                              ])
     avoid_first, avoid_last = 2, 1 # layers that are usually never important
-    vectors_to_compare = vectors_to_compare[avoid_first:-avoid_last]
-    diff_score = F.normalize(vectors_to_compare, dim=-1).diff(dim=0).norm(dim=-1)
+    assert avoid_first >= 1 # due to .diff() we will not be able to compute a score for the first layer
+    diff_score = diff_score[avoid_first-1 : len(diff_score)-avoid_last]
     important_idxs = avoid_first + diff_score.topk(k=int(np.ceil(0.1 * len(generation_texts)))).indices.cpu().numpy()
     
     # create GUI output