Torchlight provides an light-weight PyTorch trainer, as well as many useful utils, including network components, transforms, metrics, etal. for fast prototyping of computer vision models.

✨ All top level packages are self-contained and independent. Feel free to steal any part into your own project.

• Install via PyPI.

pip install torchlights

• Install the latest version from source.

git clone https://github.com/Zeqiang-Lai/torchlight.git
cd torchlight
pip install .
pip install -e . # editable installation

# or simply if you don't need editable installation
pip install git+https://github.com/Zeqiang-Lai/torchlight.git

• Most modules are self-contained.
• Debug Mode.
• User friendly progress bar .
• Save latest checkpoint if interrupted by Ctrl-C.
• Override any option in configuration file with cmd args.

• kornia: Open Source Differentiable Computer Vision Library.
• huggingface/datasets: original design for NLP, but also include some vision datasets.
• huggingface/accelerate: A simple way to train and use PyTorch models with multi-GPU, TPU, mixed-precision.
• einops: Flexible and powerful tensor operations for readable and reliable code.
• torch-fidelity: High-fidelity performance metrics for generative models in PyTorch.
• piq: Measures and metrics for image2image tasks. PyTorch.
• image-similarity-measures: Numpy implementation of eight evaluation metrics to access the similarity between two images. The eight metrics are as follows: RMSE, PSNR, SSIM, ISSM, FSIM, SRE, SAM, and UIQ.
• ResizeRight: The correct way to resize images or tensors. For Numpy or Pytorch (differentiable).
• omegaconf: Flexible Python configuration system. The last one you will ever need.