UniPercept

This package contains a collection of libraries for rapid testing and development in computer vision research settings.

Note that this is explicitly not a 'framework', but rather a collection of common utlities and boilerplate code. The individual libraries can be individually used and combined, without requiring a large shift in paradigm or major cross-dependencies.

Additionally, unipercept includes Hydra-based configurataion system and CLI to speed up common research tasks.

Installation

This package requires at least Python 3.11 and PyTorch 2.1. Once you have created an environment with these dependencies, we can proceed to install unipercept using one of three installation methods.

Stable release (recommended)

You can install the latest stable release from PyPI via

pip install unipercept

Master branch

To install the latest version, which is not guaranteed to be stable, install from GitHub using

pip install git+https://github.com/kurt-stolle/unipercept.git

Developers

If your use-case requires changes to our codebase, we recommend that you first fork this repository and download your own fork locally. Assuming you have the GitHub CLI installed, you can clone your fork with

gh repo clone unipercept

Then, you can proceed to install the package in editable mode by running

pip install --editable unipercept

You are invited to share your improvements to the codebase through a pull request on this repository. Before making a pull request, please ensure your changes follow our code guidelines by running pre-commit before adding your files into a Git commit.

Training and evaluation

Models can be trained and evalurated from the CLI or through the Python API.

CLI

To train a model with the CLI:

unipercept train --config <config path> <overrides>

Without a <config name>, an interactive prompt will be started to assist in finding a configuration file.

To evaluate a model with the CLI:

unipercept evaluate --config <config path> <overrides>

Developer guidelines

All tests can ran via python -m pytest. However, we also provide a make directive that uses pytorch-xdist to speed up the process:

make test

You may need to tune the parameters if memory problems arise during testing.

Similarly, benchmarks are implemented using pytest-benchmark. To run them, use:

make benchmark

Coverage reports are built using pytest-cov, and executed using:

make coverage

Last, we largely follow the same principles and methods as Transformers uses for testing. For more information on using pytest for automated testing, check out their documentation.

F.A.Q.

Building fais in Conda environemnts that ship their own GCC/G++ versions

If you are using a Conda environment that ships its own GCC/G++ versions, you may encounter issues when building the torch extensions shipped with unipercept. To resolve this, you can set the CC and CXX environment variables to the active environment's GCC/G++ versions before running the installation command.

In your Conda environment, run the following commands:

conda env config vars set CC=$(which gcc)
conda env config vars set CXX=$(which g++)

Then, reload the environment and install unipercept as usual.

License

This repository is released under the MIT License. For more information, please refer to the LICENSE file.