Update app.py

app.py

@@ -5,35 +5,31 @@ import torch
 from transformers import AutoImageProcessor, AutoModelForImageClassification
 
 # --- Configuration ---
-MODEL_NAME = "nateraw/fer-2013"
-# CRITICAL FIX: Explicitly and unconditionally set DEVICE to 'cpu' 
-DEVICE = "cpu" 
+# NEW, more stable ViT-based model for emotion detection
+MODEL_NAME = "abhilash88/face-emotion-detection" 
+DEVICE = "cpu" # Explicitly set to CPU
 
 # --- Model and Processor Loading ---
 try:
     processor = AutoImageProcessor.from_pretrained(MODEL_NAME)
     
-    # CRITICAL FIX: Use map_location='cpu' for memory-safe loading.
-    # We also use the local_files_only=False default to re-attempt download if necessary.
+    # Load model with map_location='cpu' for memory-safe loading.
     model = AutoModelForImageClassification.from_pretrained(
         MODEL_NAME, 
         map_location=DEVICE 
-    )
-    # Ensure all layers are formally moved to the CPU
-    model.to(DEVICE)
-    model.eval()
+    ).to(DEVICE)
     
+    model.eval()
     LABELS = model.config.id2label
     
     print(f"Model loaded successfully on device: {DEVICE}")
 
 except Exception as e:
-    # If loading fails, return a highly specific error message to the user.
     print(f"CRITICAL ERROR during model loading: {e}")
     processor = None
     model = None
-    # This error message directs the user to the only remaining physical solution
-    LABELS = {0: "System Error: Failed to load model. Likely OOM. Try a paid CPU/GPU tier."}
+    # If this ViT model fails, the only remaining cause is a lack of RAM.
+    LABELS = {0: "HARDWARE FAILURE: Free tier lacks sufficient RAM (OOM). Upgrade required."}
 
 # --- Inference Function ---
 def classify_emotion(image_np: np.ndarray) -> str:
@@ -42,24 +38,18 @@ def classify_emotion(image_np: np.ndarray) -> str:
         return LABELS[0]
 
     try:
-        # 1. Convert numpy array to PIL Image
         image = Image.fromarray(image_np).convert("RGB")
-        
-        # 2. Preprocess
         inputs = processor(images=image, return_tensors="pt").to(DEVICE)
         
-        # 3. Run inference
         with torch.no_grad():
             outputs = model(**inputs)
             
-        # 4. Process predictions
         probabilities = torch.nn.functional.softmax(outputs.logits, dim=-1)
         confidence, predicted_class_idx = torch.max(probabilities, 1)
         
         dominant_emotion = LABELS[predicted_class_idx.item()]
         confidence_score = confidence.item()
 
-        # 5. Format the result
         result_str = (
             f"<h2 class='text-xl font-bold'>Predicted Emotion:</h2>"
             f"<p class='text-3xl mt-2'>**{dominant_emotion.upper()}**</p>"
@@ -78,9 +68,9 @@ iface = gr.Interface(
         label="Upload an image of a face"
     ),
     outputs=gr.Markdown(label="Predicted Emotion"),
-    title="😊 PyTorch Facial Emotion Detection (Optimized)",
+    title="😊 PyTorch Facial Emotion Detection (ViT Model)",
     description=(
-        "Upload an image containing a face to classify the dominant emotion. Built for maximum stability on Hugging Face Spaces."
+        "Uses a stable ViT (Vision Transformer) model fine-tuned on the FER-2013 dataset."
     ),
     allow_flagging="never",
     theme=gr.themes.Soft()