models/vit.py

from timm.models.vision_transformer import VisionTransformer
import torch
from torchvision import transforms
import cv2

def load_vit(weight_path="models/weights/cvit2_deepfake_detection_ep_50.pth"):
    """
    Loads a custom Vision Transformer model that matches the checkpoint structure.
    """
    model = VisionTransformer(
        img_size=224,
        patch_size=16,
        embed_dim=1024,
        depth=12,
        num_heads=16,
        num_classes=1,
    )
    checkpoint = torch.load(weight_path, map_location="cpu")
    state_dict = checkpoint.get("state_dict", checkpoint)
    model.load_state_dict(state_dict, strict=False)
    model.eval()
    return model

def process_video(video_path, model):
    """
    Processes a video file and predicts deepfake probabilities for each frame.
    """
    transform = transforms.Compose([
        transforms.Resize((224, 224)),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ])
    
    cap = cv2.VideoCapture(video_path)
    frame_probs = []

    while cap.isOpened():
        ret, frame = cap.read()
        if not ret:
            break
        
        # Convert frame to RGB and preprocess
        frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
        input_tensor = transform(frame_rgb).unsqueeze(0)
        
        # Predict
        with torch.no_grad():
            prob = torch.sigmoid(model(input_tensor)).item()
        frame_probs.append(prob)
    
    cap.release()
    return frame_probs