data_loader.py

from huggingface_hub import hf_hub_download
import h5py
import torch
import numpy as np
from torch.utils.data import Dataset, DataLoader

def download_and_load_dataset(repo_id, filename="ucf_crime_features_labeled.h5"):
    """Download the HDF5 file from Hugging Face and return the local path."""
    hdf5_path = hf_hub_download(
        repo_id=repo_id,
        filename=filename,
        repo_type="dataset"
    )
    return hdf5_path

class UCFCrimeDataset(Dataset):
    """Dataset class for loading UCF-Crime features with temporal annotations."""
    
    def __init__(self, hdf5_path, split=None, transform=None):
        self.hdf5_path = hdf5_path
        self.transform = transform
        
        # Open the HDF5 file
        self.hdf5_file = h5py.File(hdf5_path, 'r')
        
        # Build list of video paths (category/video_name)
        self.video_paths = []
        for category_name in self.hdf5_file.keys():
            category_group = self.hdf5_file[category_name]
            for video_name in category_group.keys():
                video_path = f"{category_name}/{video_name}"
                self.video_paths.append(video_path)
        
        # Filter by split if specified
        if split is not None:
            filtered_paths = []
            for video_path in self.video_paths:
                video_group = self.hdf5_file[video_path]
                video_split = video_group.attrs.get('split', 'Unknown')
                
                # Handle bytes type
                if isinstance(video_split, bytes):
                    video_split = video_split.decode('utf-8')
                
                # Case-insensitive comparison
                if video_split.lower() == split.lower():
                    filtered_paths.append(video_path)
            
            self.video_paths = filtered_paths
            print(f"Loaded {len(self.video_paths)} videos for split: {split}")
    
    def __len__(self):
        return len(self.video_paths)
    
    def __getitem__(self, idx):
        video_path = self.video_paths[idx]
        video_group = self.hdf5_file[video_path]
        
        features = np.array(video_group['features'])
        labels = np.array(video_group['labels'])
        
        # Convert to tensors
        features = torch.from_numpy(features).float()
        labels = torch.from_numpy(labels).float()
        
        if self.transform:
            features = self.transform(features)
        
        # Get metadata
        duration = video_group.attrs.get('duration', 0.0)
        split = video_group.attrs.get('split', 'Unknown')
        if isinstance(split, bytes):
            split = split.decode('utf-8')
        
        return {
            'video_id': video_path,
            'features': features,
            'labels': labels,
            'duration': duration,
            'split': split
        }
    
    def close(self):
        """Close the HDF5 file."""
        if self.hdf5_file:
            self.hdf5_file.close()

def create_dataloaders_from_huggingface(repo_id, batch_size=16, num_workers=2):
    """Download dataset from Hugging Face and create dataloaders."""
    
    # Download the HDF5 file
    print(f"Downloading dataset from {repo_id}...")
    hdf5_path = download_and_load_dataset(repo_id)
    print(f"✓ Dataset downloaded to: {hdf5_path}")
    
    # Create datasets
    print("\nCreating datasets...")
    train_dataset = UCFCrimeDataset(hdf5_path, split='Train')
    val_dataset = UCFCrimeDataset(hdf5_path, split='Val')
    test_dataset = UCFCrimeDataset(hdf5_path, split='Test')
    
    # Create dataloaders
    train_loader = DataLoader(
        train_dataset, 
        batch_size=batch_size, 
        shuffle=True, 
        num_workers=num_workers,
        pin_memory=True
    )
    val_loader = DataLoader(
        val_dataset, 
        batch_size=batch_size, 
        shuffle=False, 
        num_workers=num_workers,
        pin_memory=True
    )
    test_loader = DataLoader(
        test_dataset, 
        batch_size=batch_size, 
        shuffle=False, 
        num_workers=num_workers,
        pin_memory=True
    )
    
    print(f"\n{'='*60}")
    print("Dataset Statistics:")
    print(f"{'='*60}")
    print(f"  Training set:   {len(train_dataset):>4} videos")
    print(f"  Validation set: {len(val_dataset):>4} videos")
    print(f"  Test set:       {len(test_dataset):>4} videos")
    print(f"  Total:          {len(train_dataset) + len(val_dataset) + len(test_dataset):>4} videos")
    print(f"{'='*60}")
    
    return train_loader, val_loader, test_loader, hdf5_path

if __name__ == "__main__":
    repo_id = "Rahima411/ucf-anomaly-detection-mapped"
    
    # Create dataloaders
    train_loader, val_loader, test_loader, hdf5_path = create_dataloaders_from_huggingface(
        repo_id=repo_id,
        batch_size=16,
        num_workers=2
    )
    
    # Test loading batches
    print("\nLoading Data...")
    print("-" * 60)
    
    for split_name, loader in [("Train", train_loader), ("Val", val_loader), ("Test", test_loader)]:
        print(f"\n{split_name} set - First batch:")
        for batch in loader:
            print(f"  Batch size: {len(batch['video_id'])} videos")
            print(f"  Features shape: {batch['features'].shape}")
            print(f"  Labels shape: {batch['labels'].shape}")
            print(f"  Sample video IDs: {batch['video_id'][:3]}")
            
            # Calculate anomaly statistics
            labels_np = batch['labels'].numpy()
            num_anomaly_frames = (labels_np == 1).sum()
            total_frames = labels_np.size
            anomaly_pct = 100 * num_anomaly_frames / total_frames if total_frames > 0 else 0
            
            print(f"  Anomaly frames: {num_anomaly_frames:,} / {total_frames:,} ({anomaly_pct:.2f}%)")
            break  # Only show first batch