Update app.py

app.py

@@ -9,6 +9,7 @@ import timm
 from torchvision import transforms
 #from llama_cpp import Llama
 from peft import PeftModel
+import traceback
 
 # 1. Model Definitions (Same as in training script)
 class SigLIPImageEncoder(torch.nn.Module):
@@ -86,6 +87,7 @@ print("phi-3 model loaded sucessfully")
 # 3. Inference Function
 
 @spaces.GPU
+# 3. Inference Function
 def predict(image, question):
     """
     Takes an image and a question as input and returns an answer.
@@ -101,26 +103,27 @@ def predict(image, question):
         with torch.no_grad():
             image_embeddings = image_encoder(image)
             # Flatten the image embeddings for simplicity
-            image_embeddings = image_embeddings.flatten().tolist()
+            image_embeddings_list = image_embeddings.flatten().tolist() # Convert to list of floats
+            image_embeddings_str = ' '.join(map(str, image_embeddings_list)) # Convert to space-separated string
 
         # Create the prompt with image embeddings
-        prompt = f"Question: {question}\nImage Embeddings: {image_embeddings}\nAnswer:"
+        prompt = f"Question: {question}\nImage Embeddings: {image_embeddings_str}\nAnswer:"
 
         # Generate answer using llama.cpp
-        output = model(
+        output = llm(
             prompt,
-            max_tokens=128,
+            max_tokens=200,
             stop=["Q:", "\n"],
             echo=False,
         )
-
         answer = output["choices"][0]["text"].strip()
 
         return answer
 
     except Exception as e:
+        traceback.print_exc()
         return f"An error occurred: {str(e)}"
-
+        
 # 4. Gradio Interface
 iface = gr.Interface(
     fn=predict,