app.py

import cv2
import torch
import numpy as np
from PIL import Image
import torchvision.transforms as transforms
import time
import os
import json
from typing import Dict, List, Any
from fastapi import FastAPI, UploadFile, File, HTTPException
from fastapi.responses import JSONResponse, HTMLResponse
import uuid
from pathlib import Path
import gradio as gr
import tempfile

app = FastAPI()

# Global variable to store the history of largest face detections
largest_face_detections = []

# EmotionCNN model definition
class EmotionCNN(torch.nn.Module):
    def __init__(self, num_classes=7):
        super(EmotionCNN, self).__init__()
        
        # First convolutional block
        self.conv1 = torch.nn.Sequential(
            torch.nn.Conv2d(1, 64, kernel_size=3, padding=1),
            torch.nn.BatchNorm2d(64),
            torch.nn.ReLU(),
            torch.nn.MaxPool2d(kernel_size=2, stride=2)
        )
        
        # Second convolutional block
        self.conv2 = torch.nn.Sequential(
            torch.nn.Conv2d(64, 128, kernel_size=3, padding=1),
            torch.nn.BatchNorm2d(128),
            torch.nn.ReLU(),
            torch.nn.MaxPool2d(kernel_size=2, stride=2)
        )
        
        # Third convolutional block
        self.conv3 = torch.nn.Sequential(
            torch.nn.Conv2d(128, 256, kernel_size=3, padding=1),
            torch.nn.BatchNorm2d(256),
            torch.nn.ReLU(),
            torch.nn.MaxPool2d(kernel_size=2, stride=2)
        )
        
        # Fourth convolutional block
        self.conv4 = torch.nn.Sequential(
            torch.nn.Conv2d(256, 512, kernel_size=3, padding=1),
            torch.nn.BatchNorm2d(512),
            torch.nn.ReLU(),
            torch.nn.MaxPool2d(kernel_size=2, stride=2)
        )
        
        # Fifth convolutional block with residual connection
        self.conv5 = torch.nn.Sequential(
            torch.nn.Conv2d(512, 512, kernel_size=3, padding=1),
            torch.nn.BatchNorm2d(512),
            torch.nn.ReLU()
        )
        
        # Attention mechanism
        self.attention = torch.nn.Sequential(
            torch.nn.Conv2d(512, 1, kernel_size=1),
            torch.nn.Sigmoid()
        )
        
        # Fully connected layers
        self.fc = torch.nn.Sequential(
            torch.nn.Dropout(0.5),
            torch.nn.Linear(512 * 3 * 3, 1024),
            torch.nn.ReLU(),
            torch.nn.Dropout(0.5),
            torch.nn.Linear(1024, 512),
            torch.nn.ReLU(),
            torch.nn.Dropout(0.3),
            torch.nn.Linear(512, num_classes)
        )
        
    def forward(self, x):
        x = self.conv1(x)
        x = self.conv2(x)
        x = self.conv3(x)
        x = self.conv4(x)
        
        # Fifth conv block with residual connection
        x_res = x
        x = self.conv5(x)
        x = x + x_res
        
        # Apply attention
        attn = self.attention(x)
        x = x * attn
        
        # Flatten
        x = x.view(x.size(0), -1)
        
        # Fully connected
        x = self.fc(x)
        return x

def load_emotion_model(model_path, device='cuda' if torch.cuda.is_available() else 'cpu'):
    """Load the emotion recognition model"""
    checkpoint = torch.load(model_path, map_location=device)
    
    model = EmotionCNN(num_classes=7)
    model.load_state_dict(checkpoint['model_state_dict'])
    model.to(device)
    model.eval()
    
    return model

def preprocess_face(face_img, size=(48, 48)):
    """Preprocess face image for emotion detection"""
    transform = transforms.Compose([
        transforms.Resize(size),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.5], std=[0.5])
    ])
    
    # Convert to PIL Image
    if isinstance(face_img, np.ndarray):
        face_img = Image.fromarray(cv2.cvtColor(face_img, cv2.COLOR_BGR2RGB))
    
    # Convert to grayscale
    face_img = face_img.convert('L')
    
    # Apply transformations
    face_tensor = transform(face_img).unsqueeze(0)
    return face_tensor

def process_video(video_path: str) -> Dict[str, Any]:
    """
    Process a video file and return emotion detection results.
    
    Args:
        video_path (str): Path to the video file
        
    Returns:
        Dict containing:
        - success (bool): Whether processing was successful
        - message (str): Status message
        - results (List[Dict]): List of emotion detection results
        - error (str): Error message if any
    """
    global largest_face_detections
    largest_face_detections = []  # Reset detections for new video
    
    # Paths - adjust these paths according to your Hugging Face Space
    face_cascade_path = cv2.data.haarcascades + 'haarcascade_frontalface_default.xml'
    emotion_model_path = "/data/best_emotion_model.pth"  # Path in Hugging Face Space
    
    # Check if models exist
    if not os.path.exists(face_cascade_path):
        return {
            "success": False,
            "message": "Face cascade classifier not found",
            "results": [],
            "error": f"Error: Face cascade classifier not found at {face_cascade_path}"
        }
        
    if not os.path.exists(emotion_model_path):
        return {
            "success": False,
            "message": "Emotion model not found",
            "results": [],
            "error": f"Error: Emotion model not found at {emotion_model_path}"
        }
    
    # Set device
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    
    # Load models
    try:
        face_cascade = cv2.CascadeClassifier(face_cascade_path)
        emotion_model = load_emotion_model(emotion_model_path, device)
    except Exception as e:
        return {
            "success": False,
            "message": "Error loading models",
            "results": [],
            "error": str(e)
        }
        
    # Emotion labels
    emotions = ['Angry', 'Disgust', 'Fear', 'Happy', 'Sad', 'Surprise', 'Neutral']
    
    # Open video
    cap = cv2.VideoCapture(video_path)
    if not cap.isOpened():
        return {
            "success": False,
            "message": "Could not open video file",
            "results": [],
            "error": f"Error: Could not open video file at {video_path}"
        }
    
    frame_count = 0
    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    
    while True:
        ret, frame = cap.read()
        if not ret:
            break
            
        frame_count += 1
        
        # Variables to track largest face
        largest_face_area = 0
        current_detection = None
        
        # Convert frame to grayscale for face detection
        gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
        
        # Detect faces using Haar Cascade
        faces = face_cascade.detectMultiScale(
            gray,
            scaleFactor=1.1,
            minNeighbors=5,
            minSize=(30, 30)
        )
        
        # Process each detected face
        for (x, y, w, h) in faces:
            # Calculate face area
            face_area = w * h
            
            # Extract face region with margin
            margin = 20
            x1 = max(0, x - margin)
            y1 = max(0, y - margin)
            x2 = min(frame.shape[1], x + w + margin)
            y2 = min(frame.shape[0], y + h + margin)
            
            face_img = frame[y1:y2, x1:x2]
            
            # Skip if face is too small
            if face_img.size == 0 or face_img.shape[0] < 20 or face_img.shape[1] < 20:
                continue
            
            # Convert face to PIL Image and preprocess
            face_tensor = preprocess_face(face_img)
            
            # Predict emotion
            with torch.no_grad():
                face_tensor = face_tensor.to(device)
                output = emotion_model(face_tensor)
                probabilities = torch.nn.functional.softmax(output, dim=1)
                emotion_idx = torch.argmax(output, dim=1).item()
                confidence = probabilities[0][emotion_idx].item()
              
            # Get emotion label
            emotion = emotions[emotion_idx]
            
            # Update largest face if current face is larger
            if face_area > largest_face_area:
                largest_face_area = face_area
                current_detection = {
                    'emotion': emotion,
                    'confidence': confidence,
                    'timestamp': time.time(),
                    'frame_number': frame_count
                }
        
        # Add current detection to history if a face was detected
        if current_detection:
            largest_face_detections.append(current_detection)
    
    # Release resources
    cap.release()
    
    # Process results
    if not largest_face_detections:
        return {
            "success": True,
            "message": "No faces detected in video",
            "results": [],
            "error": None
        }
    
    # Calculate summary statistics
    emotions_count = {}
    for detection in largest_face_detections:
        emotion = detection['emotion']
        emotions_count[emotion] = emotions_count.get(emotion, 0) + 1
    
    # Get dominant emotion
    dominant_emotion = max(emotions_count.items(), key=lambda x: x[1])[0]
    
    return {
        "success": True,
        "message": "Video processed successfully",
        "results": {
            "detections": largest_face_detections,
            "summary": {
                "total_frames": total_frames,
                "total_detections": len(largest_face_detections),
                "emotions_count": emotions_count,
                "dominant_emotion": dominant_emotion
            }
        },
        "error": None
    }

# Gradio Interface Functions
def gradio_analyze_video(video_path: str):
    """Wrapper function for Gradio interface"""
    result = process_video(video_path)
    if not result["success"]:
        return {"error": result.get("error", "Processing failed")}
    
    # Format results for better Gradio display
    summary = result["results"]["summary"]
    detections = result["results"]["detections"]
    
    output = {
        "summary": {
            "total_frames": summary["total_frames"],
            "faces_detected": summary["total_detections"],
            "dominant_emotion": summary["dominant_emotion"],
            "emotion_distribution": summary["emotions_count"]
        },
        "sample_detections": detections[:5]  # Show first 5 detections
    }
    return output

def save_upload_file_tmp(upload_file: UploadFile) -> str:
    """Save uploaded file to temporary location"""
    try:
        suffix = Path(upload_file.filename).suffix
        with tempfile.NamedTemporaryFile(delete=False, suffix=suffix) as tmp:
            tmp.write(upload_file.file.read())
            return tmp.name
    finally:
        upload_file.file.close()

# Gradio Interface
with gr.Blocks(title="Video Emotion Detection", theme=gr.themes.Soft()) as demo:
    gr.Markdown("""
    # 🎭 Video Emotion Detection
    Upload a video to analyze facial emotions frame by frame
    """)
    
    with gr.Row():
        with gr.Column():
            video_input = gr.Video(
                label="Upload Video", 
                sources=["upload"],
                type="filepath"
            )
            submit_btn = gr.Button("Analyze Video", variant="primary")
            
        with gr.Column():
            output_json = gr.JSON(label="Analysis Results")
            gr.Markdown("""
            ### Results Interpretation
            - **Dominant Emotion**: Most frequently detected emotion
            - **Emotion Distribution**: Count of each emotion detected
            - **Sample Detections**: First 5 emotion detections
            """)
    
    submit_btn.click(
        fn=gradio_analyze_video,
        inputs=video_input,
        outputs=output_json,
        api_name="predict"
    )

# FastAPI Endpoints
@app.post("/api/analyze-video")
async def analyze_video(file: UploadFile = File(...)):
    """Original FastAPI endpoint"""
    try:
        temp_path = save_upload_file_tmp(file)
        result = process_video(temp_path)
        os.unlink(temp_path)
        
        if not result["success"]:
            raise HTTPException(status_code=400, detail=result.get("error", "Processing failed"))
        return JSONResponse(content=result)
    
    except Exception as e:
        if 'temp_path' in locals() and os.path.exists(temp_path):
            os.unlink(temp_path)
        raise HTTPException(status_code=500, detail=str(e))

@app.get("/", response_class=HTMLResponse)
async def root():
    """Redirect root to Gradio interface"""
    return """
    <html>
        <head>
            <title>Video Emotion Detection</title>
            <meta http-equiv="refresh" content="0; url=/gradio" />
        </head>
        <body>
            <p>Redirecting to Gradio interface... <a href="/gradio">Click here</a> if not redirected.</p>
        </body>
    </html>
    """

# Mount Gradio app to FastAPI
app = gr.mount_gradio_app(app, demo, path="/gradio")

if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=7860)