Gradient Mechanics provides utilities for GPU accelerated loading and
preprocessing of images and videos. It combines NVIDIA's NVImageCodec,
PyNvVideoCodec and CV-CUDA libraries so that decoding and transforms happen
directly on the GPU. See the documentation in the docs/ directory for more
details on how the project is structured and how to get started.
Create and sync the development environment using uv:
uv venv
source .venv/bin/activate
uv syncRun the tests to make sure everything is set up correctly:
pytestuv run python src/tests/benchmark_huggingface_images.pyuv run python src/tests/benchmark_video_dataset.py data/humming_bird_1s.mp4ffmpeg -i data/humming_bird_1s.mp4 -c:v libx264 -x264-params keyint=1:scenecut=0 -crf 18 data/humming_bird_1s_keyframes_only.mp4ffprobe -loglevel error -select_streams v:0 -show_entries frame=pict_type -of default=noprint_wrappers=1 data/humming_bird_1s_keyframes_only.mp4Terminology:
- I-frame is a key frame. Decoding a keyframe is not dependent on other frames.
- P-frame is a predicted frame. It is dependent on an preceding frame.
- B-frame is a bi-directional predicted frame. It depends on a preceding and a succeeding frame.