attention weights done - working on cross attention weights

notvenky · Mar 4, 2024 · b462113 · b462113
1 parent a9eba24
commit b462113
Showing 1 changed file with 15 additions and 11 deletions.
diff --git a/basic_demo/hf_rep.py b/basic_demo/hf_rep.py
@@ -31,7 +31,7 @@
 ).to(DEVICE).eval()
 
 image_path = "../../Downloads/vlm/IMG_0864.jpg"
-command_list_txt = ["What is the distance between tennis ball and bottle of disinfectant", "Where is the tennis ball and bottle of disinfectant"]
+command_list_txt = ["What is the distance between tennis ball and bottle of disinfectant"] #, "Where is the tennis ball and bottle of disinfectant"]
 
 image = Image.open(image_path).convert('RGB')
 history = []
@@ -51,15 +51,21 @@
     gen_kwargs = {"max_length": 2048,
                     "do_sample": False} # "temperature": 0.9
     with torch.no_grad():
-        model_outputs = model(**inputs, output_hidden_states=True, output_attentions=True)
+        model_outputs = model(**inputs, output_hidden_states=True, output_attentions=True, extract_intermediate_representation=True, return_dict=True)
         encoder_hidden_states = model_outputs.hidden_states
-        print(dir(model_outputs)) 
+        # intermediate_representation = model_outputs.intermediate_representations
+        # print(dir(model_outputs))
         print(f"Hidden states shape: {encoder_hidden_states[-1].shape}")
-        if model_outputs.attentions is not None:
-            encoder_attentions = model_outputs.attentions
-            print(f"Attention shape: {encoder_attentions[-1].shape}")
-        else:
-            print("Attention weights are not available.")
+        intermediate_representation = model_outputs.intermediate_representations
+        print(intermediate_representation)
+
+
+
+        # if model_outputs.attentions is not None:
+        #     encoder_attentions = model_outputs.attentions
+        #     print(f"Attention shape: {encoder_attentions[-1].shape}")
+        # else:
+        #     print("Attention weights are not available.")
         intr_outputs = model.generate(**inputs, **gen_kwargs)
         # print("Methods in Model:", dir(model))
         test_outputs = model.get_output_embeddings()
@@ -71,11 +77,9 @@
         response = response.split("</s>")[0]
 
         print("Query:", query)
-        # print("Image Shape:", inputs['images'][0][0].shape if inputs['images'] else None)
-        # print("Cross Image Shape:", inputs['cross_images'][0][0].shape if inputs['cross_images'] else None)
         # print("Shapes of Inputs: input_ids", inputs['input_ids'].shape, "token_type_ids", inputs['token_type_ids'].shape, "attention_mask", inputs['attention_mask'].shape)
         print("Shapes of Intrinsic Outputs:", intr_outputs.shape)
-        print("Intrinsic Output:", intr_outputs)
+        # print("Intrinsic Output:", intr_outputs)
         print("Number of Non-Zero Items:", (intr_outputs != 0).sum().item())
         print("Shape of Reply Outputs:", reply_outputs.shape)
         print("Response:", response)