app.py

import spaces
import gradio as gr
import torch
import torch.nn.functional as F
from facenet_pytorch import MTCNN, InceptionResnetV1
import os
import numpy as np
from PIL import Image as PILImage
import zipfile
import cv2
from pytorch_grad_cam import GradCAM
from pytorch_grad_cam.utils.model_targets import ClassifierOutputTarget
from pytorch_grad_cam.utils.image import show_cam_on_image
import tempfile
import matplotlib.pyplot as plt
from tqdm import tqdm
import io

with zipfile.ZipFile("examples.zip","r") as zip_ref:
    zip_ref.extractall(".")

DEVICE = 'cuda:0' if torch.cuda.is_available() else 'cpu'

mtcnn = MTCNN(
    select_largest=False,
    post_process=False,
    device=DEVICE
).to(DEVICE).eval()

model = InceptionResnetV1(
    pretrained="vggface2",
    classify=True,
    num_classes=1,
    device=DEVICE
)

checkpoint = torch.load("resnetinceptionv1_epoch_32.pth", map_location=torch.device('cpu'))
model.load_state_dict(checkpoint['model_state_dict'])
model.to(DEVICE)
model.eval()

EXAMPLES_FOLDER = 'examples'
examples_names = os.listdir(EXAMPLES_FOLDER)
examples = []
for example_name in examples_names:
    example_path = os.path.join(EXAMPLES_FOLDER, example_name)
    label = example_name.split('_')[0]
    example = {
        'path': example_path,
        'label': label
    }
    examples.append(example)
np.random.shuffle(examples) # shuffle

@spaces.GPU
def process_frame(frame, mtcnn, model, cam, targets):
    face = mtcnn(PILImage.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)))
    if face is None:
        return frame, None, None
    
    face = face.unsqueeze(0)
    face = F.interpolate(face, size=(256, 256), mode='bilinear', align_corners=False)
    
    face = face.to(DEVICE)
    face = face.to(torch.float32)
    face = face / 255.0
    face_image_to_plot = face.squeeze(0).permute(1, 2, 0).cpu().detach().numpy()

    grayscale_cam = cam(input_tensor=face, targets=targets, eigen_smooth=True)
    grayscale_cam = grayscale_cam[0, :]
    visualization = show_cam_on_image(face_image_to_plot, grayscale_cam, use_rgb=True)
    
    with torch.no_grad():
        output = torch.sigmoid(model(face).squeeze(0))
        prediction = "real" if output.item() < 0.5 else "fake"
        confidence = 1 - output.item() if prediction == "real" else output.item()
    
    return visualization, prediction, confidence

@spaces.GPU
def analyze_video(input_video: str):
    """Analyze the video for deepfake detection"""
    cap = cv2.VideoCapture(input_video)
    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    
    target_layers = [model.block8.branch1[-1]]
    cam = GradCAM(model=model, target_layers=target_layers)
    targets = [ClassifierOutputTarget(0)]
    
    frame_confidences = []
    frame_predictions = []
    
    for _ in tqdm(range(total_frames), desc="Analyzing video"):
        ret, frame = cap.read()
        if not ret:
            break
        
        _, prediction, confidence = process_frame(frame, mtcnn, model, cam, targets)
        
        if prediction is not None and confidence is not None:
            frame_confidences.append(confidence)
            frame_predictions.append(1 if prediction == "fake" else 0)
    
    cap.release()
    
    # Calculate metrics
    fake_percentage = (sum(frame_predictions) / len(frame_predictions)) * 100 if frame_predictions else 0
    avg_confidence = np.mean(frame_confidences) if frame_confidences else 0
    
    # Create graphs
    fig, (ax1, ax2) = plt.subplots(2, 1, figsize=(10, 12))
    
    # Confidence over time
    ax1.plot(frame_confidences)
    ax1.set_title("Confidence Over Time")
    ax1.set_xlabel("Frame")
    ax1.set_ylabel("Confidence")
    ax1.set_ylim(0, 1)
    
    # Prediction distribution
    ax2.hist(frame_predictions, bins=[0, 0.5, 1], rwidth=0.8)
    ax2.set_title("Distribution of Predictions")
    ax2.set_xlabel("Prediction (0: Real, 1: Fake)")
    ax2.set_ylabel("Count")
    
    # Save plot to bytes
    buf = io.BytesIO()
    plt.savefig(buf, format='png')
    buf.seek(0)
    
    # Create progress bar image
    progress_fig, progress_ax = plt.subplots(figsize=(8, 2))
    progress_ax.barh(["Fake"], [fake_percentage], color='red')
    progress_ax.barh(["Fake"], [100 - fake_percentage], left=[fake_percentage], color='green')
    progress_ax.set_xlim(0, 100)
    progress_ax.set_title("Fake Percentage")
    progress_ax.set_xlabel("Percentage")
    progress_ax.text(fake_percentage, 0, f"{fake_percentage:.1f}%", va='center', ha='left')
    
    # Save progress bar to bytes
    progress_buf = io.BytesIO()
    progress_fig.savefig(progress_buf, format='png')
    progress_buf.seek(0)
    
    return {
        "fake_percentage": fake_percentage,
        "avg_confidence": avg_confidence,
        "analysis_plot": buf,
        "progress_bar": progress_buf,
        "total_frames": total_frames,
        "processed_frames": len(frame_confidences)
    }

def format_results(results):
    return f"""
    Analysis Results:
    - Fake Percentage: {results['fake_percentage']:.2f}%
    - Average Confidence: {results['avg_confidence']:.2f}
    - Total Frames: {results['total_frames']}
    - Processed Frames: {results['processed_frames']}
    """

def analyze_and_format(input_video):
    results = analyze_video(input_video)
    text_results = format_results(results)
    
    # Convert BytesIO to PIL Images
    analysis_plot = PILImage.open(results['analysis_plot'])
    progress_bar = PILImage.open(results['progress_bar'])
    
    return text_results, analysis_plot, progress_bar

interface = gr.Interface(
    fn=analyze_and_format,
    inputs=[
        gr.Video(label="Input Video")
    ],
    outputs=[
        gr.Textbox(label="Analysis Results"),
        gr.Image(label="Analysis Plots"),
        gr.Image(label="Fake Percentage")
    ],
    title="Video Deepfake Analysis",
    description="Upload a video to analyze for potential deepfakes.",
    examples=[]
)

if __name__ == "__main__":
    interface.launch(share=True)