torchkit is a lightweight library containing PyTorch utilities useful for day-to-day research. Its main goal is to abstract away a lot of the redundant boilerplate associated with research projects like experimental configurations, logging and model checkpointing. It consists of:
torchkit.Logger |
A wrapper around Tensorboard's SummaryWriter for safe
logging of scalars, images, videos and learning rates. Supports both numpy arrays and torch Tensors.
|
torchkit.CheckpointManager |
A port of Tensorflow's checkpoint manager that automatically manages multiple checkpoints in an experimental run. |
torchkit.experiment |
A collection of methods for setting up experiment directories. |
torchkit.layers |
A set of commonly used layers in research papers not available in vanilla PyTorch like "same" and "causal" convolution and SpatialSoftArgmax.
|
torchkit.losses |
Some useful loss functions also unavailable in vanilla PyTorch like cross entropy with label smoothing and Huber loss. |
torchkit.utils |
A bunch of helper functions for config manipulation, I/O, timing, debugging, etc. |
For more details about each module, see the documentation.
To install the latest release, run:
pip install git+https://github.com/kevinzakka/torchkit.gitFor development, clone the source code and create a virtual environment for this project:
git clone https://github.com/kevinzakka/torchkit.git
cd torchkit
pip install -e .[dev]- Thanks to Karan Desai's VirTex which I used to figure out documentation-related setup for torchkit and for just being an excellent example of stellar open-source research release.
- Thanks to seals for the excellent software development practices that I've tried to emulate in this repo.
- Thanks to Brent Yi for encouraging me to use type hinting and for letting me use his awesome Bayesian filtering library's README as a template.