app.py

import gradio as gr
import numpy as np
from PIL import Image
import torch
from transformers import AutoImageProcessor, AutoModelForImageClassification

# --- Configuration ---
# 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)
    
    # Load model with map_location='cpu' for memory-safe loading.
    model = AutoModelForImageClassification.from_pretrained(
        MODEL_NAME, 
        map_location=DEVICE 
    ).to(DEVICE)
    
    model.eval()
    LABELS = model.config.id2label
    
    print(f"Model loaded successfully on device: {DEVICE}")

except Exception as e:
    print(f"CRITICAL ERROR during model loading: {e}")
    processor = None
    model = None
    # 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:
    """Performs emotion classification on an input image (numpy array)."""
    if model is None or processor is None:
        return LABELS[0]

    try:
        image = Image.fromarray(image_np).convert("RGB")
        inputs = processor(images=image, return_tensors="pt").to(DEVICE)
        
        with torch.no_grad():
            outputs = model(**inputs)
            
        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()

        result_str = (
            f"<h2 class='text-xl font-bold'>Predicted Emotion:</h2>"
            f"<p class='text-3xl mt-2'>**{dominant_emotion.upper()}**</p>"
            f"<p class='text-lg text-gray-600 mt-1'>Confidence: {confidence_score:.2f}</p>"
        )
        return result_str

    except Exception as e:
        return f"Prediction Runtime Error: {type(e).__name__} - {str(e)}"

# --- Gradio Interface ---
iface = gr.Interface(
    fn=classify_emotion,
    inputs=gr.Image(
        type="numpy", 
        label="Upload an image of a face"
    ),
    outputs=gr.Markdown(label="Predicted Emotion"),
    title="😊 PyTorch Facial Emotion Detection (ViT Model)",
    description=(
        "Uses a stable ViT (Vision Transformer) model fine-tuned on the FER-2013 dataset."
    ),
    allow_flagging="never",
    theme=gr.themes.Soft()
)

if __name__ == "__main__":
    iface.launch()