TensorFlow Video Dataset Loader

A customizable video dataset generator for TensorFlow that allows for efficient loading, preprocessing, and batching of video data for classification tasks. Ideal for research and production pipelines dealing with video-based supervised learning.

Usage

Installation

Ensure you have the required dependencies:

pip install tensorflow==2.18.0 tqdm==4.67.1 pandas==2.2.3 numpy==2.0.2 opencv-python==4.10.0.84

Directory Structure

Your dataset should be organized as follows:

dataset_root/
│
├── class_1/
│   ├── video1.mp4
│   └── video2.mp4
│
├── class_2/
│   ├── video3.mp4
│   └── video4.mp4
│
...

Each subdirectory under dataset_root represents a class, and the videos inside are samples for that class.

Example

from video_dataset_loader import VideoDatasetLoader


video_loader = VideoDatasetLoader(
    data_path='path/to/dataset_root',
    batch_size=4,
    vid_size=(128, 128),
    normalization=True,
    shuffle=True,
    shuffle_buffer_size=500,
    reshuffle_each_iteration=True,
    drop_remainder=True
)


dataset = video_loader()

for video_batch, label_batch in dataset:
    print(video_batch.shape, label_batch.shape)

Parameters

Parameter	Type	Default	Description
data_path	str	—	Path to the dataset directory
batch_size	int	—	Batch size for training
dtype	tf.dtypes.DType	tf.float32	Output dtype of the video tensors
vid_size	tuple(int, int)	None	Resize dimensions (width, height)
normalization	bool	False	Normalize pixel values to [0, 1]
shuffle	bool	True	Shuffle data before training
shuffle_buffer_size	int	1000	Buffer size for shuffling
reshuffle_each_iteration	bool	True	Reshuffle on each epoch
drop_remainder	bool	True	Drop last batch if it's smaller than batch_size

Notes

• This generator is compatible with distributed TensorFlow training.
• To ensure consistent behavior in distributed environments, set both reshuffle_each_iteration and drop_remainder to True.
• Improper use of normalization in combination with resizing or mixed precision training can cause convergence issues.

License

This work is under an MIT License.