video_saliency

Modalities:

Text

Formats:

Size:

Libraries:

Dataset card Data Studio Files Files and versions

xet

Community

video_saliency / dataloader.py

susmit600

anonymous-saltempto-submission

Duplicate from anonymous-saltempto-submission/video_saliency

09acfc8 about 1 month ago

Raw History Blame Contribute Delete

16.5 kB

#!/usr/bin/env python3
"""
PyTorch DataLoader for eye tracking dataset using Decord-GPU.
Features:
- Efficient video loading using Decord with GPU acceleration
- Per-frame fixation loading from the published CSV layout
  (data/eyetracking/video_XXXX/video_XXXX_subject_Y.csv)
- Flexible frame indexing (start, end, negative indexing)
- Automatic padding for variable number of fixations per frame
"""

import os
import numpy as np
import torch
from torch.utils.data import Dataset, DataLoader
from pathlib import Path
from typing import Tuple, Optional, List
import pandas as pd

try:
    from decord import VideoReader, gpu
except ImportError:
    raise ImportError("Please install decord: pip install decord")


class EyeTrackingDataset(Dataset):
    """
    PyTorch Dataset for Eye tracking data with Decord-GPU video loading.
    Args:
        data_dir: Path to data directory (containing videos/, eyetracking/, and the split CSVs)
        split: 'train', 'val', or 'test'
        use_gpu: Whether to use GPU for video decoding (default: True)
        gpu_id: GPU device ID (default: 0)
        subject_mode: 'combined' to pool fixations from all subjects who viewed the video,
                      or 'subject_X' (e.g. 'subject_3') to keep only one subject's stream
        cumulative_fixations: If True, load fixations from frame 0 to end_frame
                             (e.g., for frames 15-20, return fixations 0-20)
                             If False, only return fixations for requested frames (15-20)
                             Useful for temporal models that need historical context
    """

    def __init__(
        self,
        data_dir: str,
        split: str = 'train',
        use_gpu: bool = True,
        gpu_id: int = 0,
        subject_mode: str = 'combined',
        cumulative_fixations: bool = False,
        video_dir: str = None,
        csv_base_dir: str = None,
    ):
        self.data_dir = Path(data_dir)
        self.split = split
        self.use_gpu = use_gpu
        self.gpu_id = gpu_id
        self.subject_mode = subject_mode
        self.cumulative_fixations = cumulative_fixations
        self.video_dir = Path(video_dir) if video_dir else self.data_dir / 'videos'
        self.csv_base_dir = Path(csv_base_dir) if csv_base_dir else self.data_dir / 'eyetracking'

        # Get list of videos from CSV file (not from directories!)
        split_csv = self.data_dir / f"{split}_videos.csv"
        if not split_csv.exists():
            raise FileNotFoundError(f"Split CSV file not found: {split_csv}")

        split_df = pd.read_csv(split_csv)
        self.video_names = split_df['video'].tolist()

        # Store video paths for easy access
        self.video_paths = [self.video_dir / f"{name}.mp4" for name in self.video_names]

        # Create video reader context (GPU or CPU)
        if use_gpu:
            self.ctx = gpu(gpu_id)
        else:
            self.ctx = None

        print(f"Loaded {len(self.video_names)} videos from {split} split")
        print(f"Using {'GPU' if use_gpu else 'CPU'} for video decoding")
        print(f"Subject mode: {subject_mode}")

    def __len__(self) -> int:
        return len(self.video_names)

    def _get_video_reader(self, video_idx: int) -> VideoReader:
        """Create a VideoReader for the given video index."""
        video_path = self.video_paths[video_idx]

        if not video_path.exists():
            raise FileNotFoundError(f"Video not found: {video_path}")

        if self.use_gpu:
            vr = VideoReader(str(video_path), ctx=self.ctx)
        else:
            vr = VideoReader(str(video_path))

        return vr

    def _load_fixations(
        self,
        video_idx: int,
        start_frame: int,
        end_frame: int
    ) -> np.ndarray:
        """
        Load fixations for the specified frame range from the per-subject CSVs at
        ``{csv_base_dir}/{video_name}/{video_name}_subject_*.csv``.
        Empty cells in the gaze-coordinate columns are read as NaN by pandas; out-of-screen
        samples (x outside [0, 1920] or y outside [0, 1080]) are also dropped. Both kinds of
        invalid samples are excluded from the returned tensor.
        If cumulative_fixations=True, loads from frame 0 to end_frame.
        If cumulative_fixations=False, loads from start_frame to end_frame.
        Returns:
            np.ndarray of shape (num_frames, max_fixations, 2)
            Padded with -1 where there are fewer fixations
        """
        video_name = self.video_names[video_idx]
        csv_dir = self.csv_base_dir / video_name

        if not csv_dir.exists():
            raise FileNotFoundError(f"CSV directory not found: {csv_dir}")

        # Determine actual frame range to load
        if self.cumulative_fixations:
            actual_start = 0
            actual_end = end_frame
        else:
            actual_start = start_frame
            actual_end = end_frame

        # Get CSV files
        csv_files = sorted(csv_dir.glob("*.csv"))

        # Filter by subject_mode if needed
        if self.subject_mode != 'combined':
            subject_id = self.subject_mode.split('_')[-1]
            csv_files = [f for f in csv_files if f.stem.endswith(f'_{subject_id}')]

        if len(csv_files) == 0:
            # No CSV files, return empty fixations
            num_frames = actual_end - actual_start
            return np.full((num_frames, 1, 2), -1.0, dtype=np.float32)

        # Process each CSV file
        all_fixations_per_frame = {}
        RESOLUTION = (1920, 1080)

        for csv_file in csv_files:
            try:
                df = pd.read_csv(csv_file)

                df['left_x'] = df['left_x'].astype(float)
                df['left_y'] = df['left_y'].astype(float)
                df['frames'] = df['frames'].astype(float)

                # Filter invalid coordinates
                df.loc[(df['left_x'] <= 0) | (df['left_x'] >= RESOLUTION[0]), 'left_x'] = np.nan
                df.loc[(df['left_y'] <= 0) | (df['left_y'] >= RESOLUTION[1]), 'left_y'] = np.nan

                # Group by frame
                for frame_idx in range(actual_start, actual_end):
                    frame_samples = df[(df['frames'] >= frame_idx) & (df['frames'] < frame_idx + 1)]

                    for _, row in frame_samples.iterrows():
                        if not np.isnan(row['left_x']) and not np.isnan(row['left_y']):
                            if frame_idx not in all_fixations_per_frame:
                                all_fixations_per_frame[frame_idx] = []
                            all_fixations_per_frame[frame_idx].append((row['left_x'], row['left_y']))

            except Exception as e:
                print(f"Warning: Error loading CSV {csv_file}: {e}")
                continue

        # Convert to padded array
        frame_fixations = []
        max_num_fixations = 0

        for frame_idx in range(actual_start, actual_end):
            if frame_idx in all_fixations_per_frame:
                fixations = np.array(all_fixations_per_frame[frame_idx], dtype=np.float32)
                frame_fixations.append(fixations)
                max_num_fixations = max(max_num_fixations, len(fixations))
            else:
                frame_fixations.append(np.array([]).reshape(0, 2))

        # Ensure at least 1 fixation slot
        if max_num_fixations == 0:
            max_num_fixations = 1

        # Pad to max fixations
        num_frames = len(frame_fixations)
        padded_fixations = np.full((num_frames, max_num_fixations, 2), -1.0, dtype=np.float32)

        for i, fixations in enumerate(frame_fixations):
            if len(fixations) > 0:
                padded_fixations[i, :len(fixations), :] = fixations

        return padded_fixations

    def __getitem__(
        self,
        index: Tuple[int, Optional[int], Optional[int]]
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """
        Get video frames and fixations for the specified range.
        Args:
            index: Tuple of (video_idx, start_frame, end_frame)
                - video_idx: Index of video in dataset
                - start_frame: Starting frame (None = 0, negative = from end)
                - end_frame: Ending frame (None = last frame, negative = from end)
        Returns:
            video: Tensor of shape (num_frames, channels, height, width)
            fixations: Tensor of shape (num_frames, max_fixations, 2)
                       Padded with -1 where there are no fixations
        """
        if isinstance(index, int):
            # If just an integer, return the entire video
            video_idx = index
            start_frame = None
            end_frame = None
        else:
            video_idx, start_frame, end_frame = index

        # Create video reader
        vr = self._get_video_reader(video_idx)
        total_frames = len(vr)

        # Handle None and negative indexing
        if start_frame is None:
            start_frame = 0
        elif start_frame < 0:
            start_frame = total_frames + start_frame

        if end_frame is None:
            end_frame = total_frames
        elif end_frame < 0:
            end_frame = total_frames + end_frame

        # Clamp to valid range
        start_frame = max(0, min(start_frame, total_frames - 1))
        end_frame = max(start_frame + 1, min(end_frame, total_frames))

        # Load video frames
        frame_indices = list(range(start_frame, end_frame))
        frames = vr.get_batch(frame_indices).asnumpy()  # Shape: (num_frames, H, W, C)

        # Convert to torch tensor and permute to (num_frames, C, H, W)
        video = torch.from_numpy(frames).permute(0, 3, 1, 2).float()

        # Normalize to [0, 1]
        video = video / 255.0

        # Load fixations
        fixations_np = self._load_fixations(video_idx, start_frame, end_frame)
        fixations = torch.from_numpy(fixations_np).float()

        return video, fixations

    def get_video_info(self, video_idx: int) -> dict:
        """Get metadata about a video."""
        vr = self._get_video_reader(video_idx)

        return {
            'video_name': self.video_names[video_idx],
            'num_frames': len(vr),
            'fps': vr.get_avg_fps(),
            'width': vr[0].shape[1],
            'height': vr[0].shape[0],
        }


class EyeTrackingCollator:
    """
    Custom collate function for batching videos with different lengths.
    Pads videos and fixations to the same length within a batch.
    """

    def __init__(self, padding_value: float = 0.0):
        self.padding_value = padding_value

    def __call__(self, batch: List[Tuple[torch.Tensor, torch.Tensor]]) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        """
        Collate a batch of videos and fixations.
        Args:
            batch: List of (video, fixations) tuples
        Returns:
            videos: Tensor of shape (batch_size, max_frames, C, H, W)
            fixations: Tensor of shape (batch_size, max_frames, max_fixations, 2)
            lengths: Tensor of shape (batch_size,) with actual lengths
        """
        videos = [item[0] for item in batch]
        fixations = [item[1] for item in batch]

        # Get maximum dimensions
        max_frames = max(v.shape[0] for v in videos)
        max_fixations = max(f.shape[1] for f in fixations)

        # Get video dimensions (assuming all videos have same H, W, C)
        C, H, W = videos[0].shape[1], videos[0].shape[2], videos[0].shape[3]

        batch_size = len(videos)

        # Pad videos
        padded_videos = torch.full(
            (batch_size, max_frames, C, H, W),
            self.padding_value,
            dtype=videos[0].dtype
        )

        # Pad fixations
        padded_fixations = torch.full(
            (batch_size, max_frames, max_fixations, 2),
            -1.0,
            dtype=fixations[0].dtype
        )

        # Store actual lengths
        lengths = torch.zeros(batch_size, dtype=torch.long)

        for i, (video, fix) in enumerate(zip(videos, fixations)):
            num_frames = video.shape[0]
            num_fix = fix.shape[1]

            padded_videos[i, :num_frames] = video
            padded_fixations[i, :num_frames, :num_fix] = fix
            lengths[i] = num_frames

        return padded_videos, padded_fixations, lengths


def create_dataloader(
    data_dir: str,
    split: str = 'train',
    batch_size: int = 1,
    use_gpu: bool = True,
    gpu_id: int = 0,
    subject_mode: str = 'combined',
    num_workers: int = 4,
    shuffle: bool = True,
    collate_fn: Optional[callable] = None
) -> DataLoader:
    """
    Create a DataLoader for the Eye tracking dataset.
    Args:
        data_dir: Path to transformed_data directory
        split: 'train', 'val', or 'test'
        batch_size: Batch size (default: 1)
        use_gpu: Whether to use GPU for video decoding
        gpu_id: GPU device ID
        subject_mode: 'combined' or 'subject_X' for per-subject fixations
        num_workers: Number of worker processes
        shuffle: Whether to shuffle data
        collate_fn: Custom collate function (default: EyeTrackingCollator)
    Returns:
        DataLoader instance
    """
    dataset = EyeTrackingDataset(
        data_dir=data_dir,
        split=split,
        use_gpu=use_gpu,
        gpu_id=gpu_id,
        subject_mode=subject_mode
    )

    if collate_fn is None:
        collate_fn = EyeTrackingCollator()

    dataloader = DataLoader(
        dataset,
        batch_size=batch_size,
        shuffle=shuffle,
        num_workers=num_workers,
        collate_fn=collate_fn,
        pin_memory=True if use_gpu else False
    )

    return dataloader


if __name__ == "__main__":
    # Example usage
    print("=" * 80)
    print("Eye Tracking Dataset - Example Usage")
    print("=" * 80)

    data_dir = Path(__file__).parent / "data"

    # Create dataset
    dataset = EyeTrackingDataset(
        data_dir=data_dir,
        split='train',
        use_gpu=False,
        gpu_id=0,
        subject_mode='combined'
    )

    print(f"\nDataset size: {len(dataset)}")

    # Get info about first video
    info = dataset.get_video_info(0)
    print(f"\nFirst video info:")
    for key, value in info.items():
        print(f"  {key}: {value}")

    # Example 1: Load entire video
    print("\n" + "=" * 80)
    print("Example 1: Load entire video")
    print("=" * 80)
    video, fixations = dataset[0]
    print(f"Video shape: {video.shape}")  # (num_frames, C, H, W)
    print(f"Fixations shape: {fixations.shape}")  # (num_frames, max_fixations, 2)
    print(f"Video dtype: {video.dtype}, range: [{video.min():.3f}, {video.max():.3f}]")

    # Example 2: Load frame range (frames 0-30)
    print("\n" + "=" * 80)
    print("Example 2: Load frames 10-40")
    print("=" * 80)
    video, fixations = dataset[(0, 10, 40)]
    print(f"Video shape: {video.shape}")
    print(f"Fixations shape: {fixations.shape}")

    # Example 3: Load last 50 frames (negative indexing)
    print("\n" + "=" * 80)
    print("Example 3: Load last 50 frames (negative indexing)")
    print("=" * 80)
    video, fixations = dataset[(0, -50, None)]
    print(f"Video shape: {video.shape}")
    print(f"Fixations shape: {fixations.shape}")

    # Example 4: Create DataLoader with batching
    print("\n" + "=" * 80)
    print("Example 4: DataLoader with batching")
    print("=" * 80)

    # Custom dataset that returns tuples for the collator
    class TupleDataset(EyeTrackingDataset):
        def __init__(self, *args, frame_range=(0, 30), **kwargs):
            super().__init__(*args, **kwargs)
            self.frame_range = frame_range

        def __getitem__(self, idx):
            return super().__getitem__((idx, *self.frame_range))

    tuple_dataset = TupleDataset(
        data_dir=data_dir,
        split='train',
        use_gpu=False,
        gpu_id=0,
        subject_mode='combined',
        frame_range=(10, 40)
    )

    dataloader = DataLoader(
        tuple_dataset,
        batch_size=2,
        shuffle=False,
        num_workers=0,
        collate_fn=EyeTrackingCollator()
    )

    # Get one batch
    for videos, fixations, lengths in dataloader:
        print(f"Batch videos shape: {videos.shape}")  # (batch_size, max_frames, C, H, W)
        print(f"Batch fixations shape: {fixations.shape}")  # (batch_size, max_frames, max_fixations, 2)
        print(f"Actual lengths: {lengths}")
        break

    print("\n" + "=" * 80)
    print("Examples completed successfully!")
    print("=" * 80)

	#!/usr/bin/env python3
	"""
	PyTorch DataLoader for eye tracking dataset using Decord-GPU.

	Features:
	- Efficient video loading using Decord with GPU acceleration
	- Per-frame fixation loading from the published CSV layout
	(data/eyetracking/video_XXXX/video_XXXX_subject_Y.csv)
	- Flexible frame indexing (start, end, negative indexing)
	- Automatic padding for variable number of fixations per frame
	"""

	import os
	import numpy as np
	import torch
	from torch.utils.data import Dataset, DataLoader
	from pathlib import Path
	from typing import Tuple, Optional, List
	import pandas as pd

	try:
	from decord import VideoReader, gpu
	except ImportError:
	raise ImportError("Please install decord: pip install decord")


	class EyeTrackingDataset(Dataset):
	"""
	PyTorch Dataset for Eye tracking data with Decord-GPU video loading.

	Args:
	data_dir: Path to data directory (containing videos/, eyetracking/, and the split CSVs)
	split: 'train', 'val', or 'test'
	use_gpu: Whether to use GPU for video decoding (default: True)
	gpu_id: GPU device ID (default: 0)
	subject_mode: 'combined' to pool fixations from all subjects who viewed the video,
	or 'subject_X' (e.g. 'subject_3') to keep only one subject's stream
	cumulative_fixations: If True, load fixations from frame 0 to end_frame
	(e.g., for frames 15-20, return fixations 0-20)
	If False, only return fixations for requested frames (15-20)
	Useful for temporal models that need historical context
	"""

	def __init__(
	self,
	data_dir: str,
	split: str = 'train',
	use_gpu: bool = True,
	gpu_id: int = 0,
	subject_mode: str = 'combined',
	cumulative_fixations: bool = False,
	video_dir: str = None,
	csv_base_dir: str = None,
	):
	self.data_dir = Path(data_dir)
	self.split = split
	self.use_gpu = use_gpu
	self.gpu_id = gpu_id
	self.subject_mode = subject_mode
	self.cumulative_fixations = cumulative_fixations
	self.video_dir = Path(video_dir) if video_dir else self.data_dir / 'videos'
	self.csv_base_dir = Path(csv_base_dir) if csv_base_dir else self.data_dir / 'eyetracking'

	# Get list of videos from CSV file (not from directories!)
	split_csv = self.data_dir / f"{split}_videos.csv"
	if not split_csv.exists():
	raise FileNotFoundError(f"Split CSV file not found: {split_csv}")

	split_df = pd.read_csv(split_csv)
	self.video_names = split_df['video'].tolist()

	# Store video paths for easy access
	self.video_paths = [self.video_dir / f"{name}.mp4" for name in self.video_names]

	# Create video reader context (GPU or CPU)
	if use_gpu:
	self.ctx = gpu(gpu_id)
	else:
	self.ctx = None

	print(f"Loaded {len(self.video_names)} videos from {split} split")
	print(f"Using {'GPU' if use_gpu else 'CPU'} for video decoding")
	print(f"Subject mode: {subject_mode}")

	def __len__(self) -> int:
	return len(self.video_names)

	def _get_video_reader(self, video_idx: int) -> VideoReader:
	"""Create a VideoReader for the given video index."""
	video_path = self.video_paths[video_idx]

	if not video_path.exists():
	raise FileNotFoundError(f"Video not found: {video_path}")

	if self.use_gpu:
	vr = VideoReader(str(video_path), ctx=self.ctx)
	else:
	vr = VideoReader(str(video_path))

	return vr

	def _load_fixations(
	self,
	video_idx: int,
	start_frame: int,
	end_frame: int
	) -> np.ndarray:
	"""
	Load fixations for the specified frame range from the per-subject CSVs at
	``{csv_base_dir}/{video_name}/{video_name}_subject_*.csv``.

	Empty cells in the gaze-coordinate columns are read as NaN by pandas; out-of-screen
	samples (x outside [0, 1920] or y outside [0, 1080]) are also dropped. Both kinds of
	invalid samples are excluded from the returned tensor.

	If cumulative_fixations=True, loads from frame 0 to end_frame.
	If cumulative_fixations=False, loads from start_frame to end_frame.

	Returns:
	np.ndarray of shape (num_frames, max_fixations, 2)
	Padded with -1 where there are fewer fixations
	"""
	video_name = self.video_names[video_idx]
	csv_dir = self.csv_base_dir / video_name

	if not csv_dir.exists():
	raise FileNotFoundError(f"CSV directory not found: {csv_dir}")

	# Determine actual frame range to load
	if self.cumulative_fixations:
	actual_start = 0
	actual_end = end_frame
	else:
	actual_start = start_frame
	actual_end = end_frame

	# Get CSV files
	csv_files = sorted(csv_dir.glob("*.csv"))

	# Filter by subject_mode if needed
	if self.subject_mode != 'combined':
	subject_id = self.subject_mode.split('_')[-1]
	csv_files = [f for f in csv_files if f.stem.endswith(f'_{subject_id}')]

	if len(csv_files) == 0:
	# No CSV files, return empty fixations
	num_frames = actual_end - actual_start
	return np.full((num_frames, 1, 2), -1.0, dtype=np.float32)

	# Process each CSV file
	all_fixations_per_frame = {}
	RESOLUTION = (1920, 1080)

	for csv_file in csv_files:
	try:
	df = pd.read_csv(csv_file)

	df['left_x'] = df['left_x'].astype(float)
	df['left_y'] = df['left_y'].astype(float)
	df['frames'] = df['frames'].astype(float)

	# Filter invalid coordinates
	df.loc[(df['left_x'] <= 0) \| (df['left_x'] >= RESOLUTION[0]), 'left_x'] = np.nan
	df.loc[(df['left_y'] <= 0) \| (df['left_y'] >= RESOLUTION[1]), 'left_y'] = np.nan

	# Group by frame
	for frame_idx in range(actual_start, actual_end):
	frame_samples = df[(df['frames'] >= frame_idx) & (df['frames'] < frame_idx + 1)]

	for _, row in frame_samples.iterrows():
	if not np.isnan(row['left_x']) and not np.isnan(row['left_y']):
	if frame_idx not in all_fixations_per_frame:
	all_fixations_per_frame[frame_idx] = []
	all_fixations_per_frame[frame_idx].append((row['left_x'], row['left_y']))

	except Exception as e:
	print(f"Warning: Error loading CSV {csv_file}: {e}")
	continue

	# Convert to padded array
	frame_fixations = []
	max_num_fixations = 0

	for frame_idx in range(actual_start, actual_end):
	if frame_idx in all_fixations_per_frame:
	fixations = np.array(all_fixations_per_frame[frame_idx], dtype=np.float32)
	frame_fixations.append(fixations)
	max_num_fixations = max(max_num_fixations, len(fixations))
	else:
	frame_fixations.append(np.array([]).reshape(0, 2))

	# Ensure at least 1 fixation slot
	if max_num_fixations == 0:
	max_num_fixations = 1

	# Pad to max fixations
	num_frames = len(frame_fixations)
	padded_fixations = np.full((num_frames, max_num_fixations, 2), -1.0, dtype=np.float32)

	for i, fixations in enumerate(frame_fixations):
	if len(fixations) > 0:
	padded_fixations[i, :len(fixations), :] = fixations

	return padded_fixations

	def __getitem__(
	self,
	index: Tuple[int, Optional[int], Optional[int]]
	) -> Tuple[torch.Tensor, torch.Tensor]:
	"""
	Get video frames and fixations for the specified range.

	Args:
	index: Tuple of (video_idx, start_frame, end_frame)
	- video_idx: Index of video in dataset
	- start_frame: Starting frame (None = 0, negative = from end)
	- end_frame: Ending frame (None = last frame, negative = from end)

	Returns:
	video: Tensor of shape (num_frames, channels, height, width)
	fixations: Tensor of shape (num_frames, max_fixations, 2)
	Padded with -1 where there are no fixations
	"""
	if isinstance(index, int):
	# If just an integer, return the entire video
	video_idx = index
	start_frame = None
	end_frame = None
	else:
	video_idx, start_frame, end_frame = index

	# Create video reader
	vr = self._get_video_reader(video_idx)
	total_frames = len(vr)

	# Handle None and negative indexing
	if start_frame is None:
	start_frame = 0
	elif start_frame < 0:
	start_frame = total_frames + start_frame

	if end_frame is None:
	end_frame = total_frames
	elif end_frame < 0:
	end_frame = total_frames + end_frame

	# Clamp to valid range
	start_frame = max(0, min(start_frame, total_frames - 1))
	end_frame = max(start_frame + 1, min(end_frame, total_frames))

	# Load video frames
	frame_indices = list(range(start_frame, end_frame))
	frames = vr.get_batch(frame_indices).asnumpy() # Shape: (num_frames, H, W, C)

	# Convert to torch tensor and permute to (num_frames, C, H, W)
	video = torch.from_numpy(frames).permute(0, 3, 1, 2).float()

	# Normalize to [0, 1]
	video = video / 255.0

	# Load fixations
	fixations_np = self._load_fixations(video_idx, start_frame, end_frame)
	fixations = torch.from_numpy(fixations_np).float()

	return video, fixations

	def get_video_info(self, video_idx: int) -> dict:
	"""Get metadata about a video."""
	vr = self._get_video_reader(video_idx)

	return {
	'video_name': self.video_names[video_idx],
	'num_frames': len(vr),
	'fps': vr.get_avg_fps(),
	'width': vr[0].shape[1],
	'height': vr[0].shape[0],
	}


	class EyeTrackingCollator:
	"""
	Custom collate function for batching videos with different lengths.
	Pads videos and fixations to the same length within a batch.
	"""

	def __init__(self, padding_value: float = 0.0):
	self.padding_value = padding_value

	def __call__(self, batch: List[Tuple[torch.Tensor, torch.Tensor]]) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
	"""
	Collate a batch of videos and fixations.

	Args:
	batch: List of (video, fixations) tuples

	Returns:
	videos: Tensor of shape (batch_size, max_frames, C, H, W)
	fixations: Tensor of shape (batch_size, max_frames, max_fixations, 2)
	lengths: Tensor of shape (batch_size,) with actual lengths
	"""
	videos = [item[0] for item in batch]
	fixations = [item[1] for item in batch]

	# Get maximum dimensions
	max_frames = max(v.shape[0] for v in videos)
	max_fixations = max(f.shape[1] for f in fixations)

	# Get video dimensions (assuming all videos have same H, W, C)
	C, H, W = videos[0].shape[1], videos[0].shape[2], videos[0].shape[3]

	batch_size = len(videos)

	# Pad videos
	padded_videos = torch.full(
	(batch_size, max_frames, C, H, W),
	self.padding_value,
	dtype=videos[0].dtype
	)

	# Pad fixations
	padded_fixations = torch.full(
	(batch_size, max_frames, max_fixations, 2),
	-1.0,
	dtype=fixations[0].dtype
	)

	# Store actual lengths
	lengths = torch.zeros(batch_size, dtype=torch.long)

	for i, (video, fix) in enumerate(zip(videos, fixations)):
	num_frames = video.shape[0]
	num_fix = fix.shape[1]

	padded_videos[i, :num_frames] = video
	padded_fixations[i, :num_frames, :num_fix] = fix
	lengths[i] = num_frames

	return padded_videos, padded_fixations, lengths


	def create_dataloader(
	data_dir: str,
	split: str = 'train',
	batch_size: int = 1,
	use_gpu: bool = True,
	gpu_id: int = 0,
	subject_mode: str = 'combined',
	num_workers: int = 4,
	shuffle: bool = True,
	collate_fn: Optional[callable] = None
	) -> DataLoader:
	"""
	Create a DataLoader for the Eye tracking dataset.

	Args:
	data_dir: Path to transformed_data directory
	split: 'train', 'val', or 'test'
	batch_size: Batch size (default: 1)
	use_gpu: Whether to use GPU for video decoding
	gpu_id: GPU device ID
	subject_mode: 'combined' or 'subject_X' for per-subject fixations
	num_workers: Number of worker processes
	shuffle: Whether to shuffle data
	collate_fn: Custom collate function (default: EyeTrackingCollator)

	Returns:
	DataLoader instance
	"""
	dataset = EyeTrackingDataset(
	data_dir=data_dir,
	split=split,
	use_gpu=use_gpu,
	gpu_id=gpu_id,
	subject_mode=subject_mode
	)

	if collate_fn is None:
	collate_fn = EyeTrackingCollator()

	dataloader = DataLoader(
	dataset,
	batch_size=batch_size,
	shuffle=shuffle,
	num_workers=num_workers,
	collate_fn=collate_fn,
	pin_memory=True if use_gpu else False
	)

	return dataloader


	if __name__ == "__main__":
	# Example usage
	print("=" * 80)
	print("Eye Tracking Dataset - Example Usage")
	print("=" * 80)

	data_dir = Path(__file__).parent / "data"

	# Create dataset
	dataset = EyeTrackingDataset(
	data_dir=data_dir,
	split='train',
	use_gpu=False,
	gpu_id=0,
	subject_mode='combined'
	)

	print(f"\nDataset size: {len(dataset)}")

	# Get info about first video
	info = dataset.get_video_info(0)
	print(f"\nFirst video info:")
	for key, value in info.items():
	print(f" {key}: {value}")

	# Example 1: Load entire video
	print("\n" + "=" * 80)
	print("Example 1: Load entire video")
	print("=" * 80)
	video, fixations = dataset[0]
	print(f"Video shape: {video.shape}") # (num_frames, C, H, W)
	print(f"Fixations shape: {fixations.shape}") # (num_frames, max_fixations, 2)
	print(f"Video dtype: {video.dtype}, range: [{video.min():.3f}, {video.max():.3f}]")

	# Example 2: Load frame range (frames 0-30)
	print("\n" + "=" * 80)
	print("Example 2: Load frames 10-40")
	print("=" * 80)
	video, fixations = dataset[(0, 10, 40)]
	print(f"Video shape: {video.shape}")
	print(f"Fixations shape: {fixations.shape}")

	# Example 3: Load last 50 frames (negative indexing)
	print("\n" + "=" * 80)
	print("Example 3: Load last 50 frames (negative indexing)")
	print("=" * 80)
	video, fixations = dataset[(0, -50, None)]
	print(f"Video shape: {video.shape}")
	print(f"Fixations shape: {fixations.shape}")

	# Example 4: Create DataLoader with batching
	print("\n" + "=" * 80)
	print("Example 4: DataLoader with batching")
	print("=" * 80)

	# Custom dataset that returns tuples for the collator
	class TupleDataset(EyeTrackingDataset):
	def __init__(self, args, frame_range=(0, 30), *kwargs):
	super().__init__(args, *kwargs)
	self.frame_range = frame_range

	def __getitem__(self, idx):
	return super().__getitem__((idx, *self.frame_range))

	tuple_dataset = TupleDataset(
	data_dir=data_dir,
	split='train',
	use_gpu=False,
	gpu_id=0,
	subject_mode='combined',
	frame_range=(10, 40)
	)

	dataloader = DataLoader(
	tuple_dataset,
	batch_size=2,
	shuffle=False,
	num_workers=0,
	collate_fn=EyeTrackingCollator()
	)

	# Get one batch
	for videos, fixations, lengths in dataloader:
	print(f"Batch videos shape: {videos.shape}") # (batch_size, max_frames, C, H, W)
	print(f"Batch fixations shape: {fixations.shape}") # (batch_size, max_frames, max_fixations, 2)
	print(f"Actual lengths: {lengths}")
	break

	print("\n" + "=" * 80)
	print("Examples completed successfully!")
	print("=" * 80)