fix added

SkinGPT.py

@@ -296,49 +296,14 @@ class SkinGPT4(nn.Module):
         print(f"\n[DEBUG] Before replacement:")
         print(f"Text embeddings shape: {input_embeddings.shape}")
         print(f"Visual embeddings shape: {visual_embeds.shape}")
-        print(f"Image token embedding (before):\n{input_embeddings}...")
+        replaced_pos = replace_positions[1][0]
+        print(f"Image token embedding (before):\n{input_embeddings[0, replaced_pos, :5]}...")
 
         for pos in replace_positions:
             input_embeddings[0, pos[1]] = visual_embeds[0]
 
-        # if visual_embeds.dtype != input_embeddings.dtype:
-        #     visual_embeds = visual_embeds.to(input_embeddings.dtype)
-        # input_embeddings[image_token_pos] = visual_embeds
-
         print(f"\n[DEBUG] After replacement:")
-        print(f"Image token embedding (after):\n{input_embeddings}...")
-
-        # outputs = self.llama.generate(
-        #     inputs_embeds=input_embeddings,
-        #     max_new_tokens=max_length,
-        #     temperature=0.7,
-        #     top_p=0.9,
-        #     repetition_penalty=1.2,  # Prevent repetition
-        #     do_sample=True,
-        #     pad_token_id=self.tokenizer.eos_token_id,
-        #     eos_token_id=self.tokenizer.eos_token_id
-        # )
-
-        # outputs = self.llama.generate(
-        #     inputs_embeds=input_embeddings,
-        #     max_new_tokens=max_new_tokens,
-        #     num_beams=1,
-        #     do_sample=True,
-        #     min_length=1,
-        #     top_p=0.9,
-        #     repetition_penalty=1.1,
-        #     length_penalty=1,
-        #     temperature=1.0,
-        #     pad_token_id=self.tokenizer.eos_token_id
-        # )
-
-        with torch.no_grad():
-            # Test forward pass without generation
-            test_outputs = self.llama(
-                inputs_embeds=input_embeddings,
-                output_hidden_states=True
-            )
-            print(f"\n[DEBUG] First 5 output logits:\n{test_outputs.logits[0, :5, :5]}")
+        print(f"Image token embedding (after):\n{input_embeddings[0, replaced_pos, :5]}...")
 
         outputs = self.llama.generate(
             inputs_embeds=input_embeddings,
@@ -356,7 +321,7 @@ class SkinGPT4(nn.Module):
         full_output = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
         print(f"Full Output from llama : {full_output}")
         response = full_output.split("### Response:")[-1].strip()
-        print(f"Response from llama : {full_output}")
+        # print(f"Response from llama : {full_output}")
 
         return response