Add pneumonia detection app with Grad-CAM

app.py

@@ -86,7 +86,7 @@ def image_to_base64(image):
     return f"data:image/jpeg;base64,{img_base64}"
 
 def query_medgemma(message, history, image=None):
-    """Query MedGemma endpoint with text and optional image"""
+    """Query MedGemma endpoint with proper multimodal format"""
     
     # Your endpoint URL
     endpoint_url = "https://t911ok4t5x994zcu.us-east-1.aws.endpoints.huggingface.cloud"
@@ -97,50 +97,90 @@ def query_medgemma(message, history, image=None):
         "Content-Type": "application/json"
     }
     
-    # Prepare the message content
-    content = []
-    
-    # Add image if provided
+    # Prepare payload based on whether we have an image or not
     if image is not None:
+        # Multimodal format: Send image as base64 in the content
         image_base64 = image_to_base64(image)
-        content.append({
-            "type": "image_url",
-            "image_url": {"url": image_base64}
-        })
-    
-    # Add text message
-    content.append({
-        "type": "text", 
-        "text": message
-    })
-    
-    # Prepare payload
-    payload = {
-        "model": "tgi",
-        "messages": [
-            {
-                "role": "user",
-                "content": content
+        
+        # TGI multimodal format for MedGemma
+        payload = {
+            "inputs": {
+                "text": message,
+                "image": image_base64
+            },
+            "parameters": {
+                "max_new_tokens": 500,
+                "temperature": 0.7,
+                "do_sample": True,
+                "return_full_text": False
             }
-        ],
-        "max_tokens": 500,
-        "temperature": 0.7
-    }
+        }
+        
+        # Alternative format if the above doesn't work
+        payload_alt = {
+            "inputs": message,
+            "image": image_base64,
+            "parameters": {
+                "max_new_tokens": 500,
+                "temperature": 0.7,
+                "do_sample": True,
+                "return_full_text": False
+            }
+        }
+        
+    else:
+        # Text-only format
+        payload = {
+            "inputs": message,
+            "parameters": {
+                "max_new_tokens": 500,
+                "temperature": 0.7,
+                "do_sample": True,
+                "return_full_text": False,
+                "stop": ["<|im_end|>", "</s>"]
+            }
+        }
+        payload_alt = None
     
     try:
+        # Try primary format
         response = requests.post(endpoint_url, headers=headers, json=payload, timeout=30)
         
         if response.status_code == 200:
             result = response.json()
-            if "choices" in result and len(result["choices"]) > 0:
-                return result["choices"][0]["message"]["content"]
+            
+            # Handle different TGI response formats
+            if isinstance(result, list) and len(result) > 0:
+                return result[0].get("generated_text", "Error en el formato de respuesta")
+            elif "generated_text" in result:
+                return result["generated_text"]
+            else:
+                return f"Formato de respuesta inesperado: {str(result)[:200]}"
+                
+        elif response.status_code == 422 and payload_alt is not None:
+            # Try alternative format for multimodal
+            response = requests.post(endpoint_url, headers=headers, json=payload_alt, timeout=30)
+            
+            if response.status_code == 200:
+                result = response.json()
+                if isinstance(result, list) and len(result) > 0:
+                    return result[0].get("generated_text", "Error en el formato de respuesta")
+                elif "generated_text" in result:
+                    return result["generated_text"]
+                else:
+                    return f"Formato de respuesta inesperado: {str(result)[:200]}"
             else:
-                return "Lo siento, no pude obtener una respuesta del modelo."
+                return f"Error 422 en ambos formatos. Detalles: {response.text[:300]}"
+                
+        elif response.status_code == 503:
+            return "El modelo está escalado a cero. Intenta de nuevo en unos segundos mientras se activa."
+        elif response.status_code == 422:
+            return f"Error de formato en la petición. Detalles: {response.text[:300]}"
         else:
-            return f"Error del endpoint: {response.status_code}. El modelo puede estar escalado a cero - intenta de nuevo en unos segundos."
+            return f"Error del endpoint: {response.status_code}. Detalles: {response.text[:200]}"
             
     except requests.exceptions.Timeout:
-        return "Timeout: El modelo está despertando, intenta de nuevo en unos segundos."
+        return "Timeout: El modelo está procesando, intenta de nuevo en unos segundos."
     except Exception as e:
         return f"Error de conexión: {str(e)}"