v0.0.20: Python 3.13 support

• This release adds support for Python 3.13 and drops support for Python 3.10.
• Similar to PyTorch, we also drop support for macOS 10.14+ x86-64.
• We now provide a source distribution in addition to our precompiled wheels so that you can compile the htc package yourself on your system against a custom (and existing) torch installation (see installation instructions for details).

v0.0.19: PyTorch compatibility and model upgrades

This release is compatible with PyTorch 2.6. This PyTorch release contains breaking changes when it comes to model loading. If a model checkpoint included more than just the weights of the model (default option for ModelCheckpoint), then this fails now with torch.load without setting weights_only=False explicitly (including all the security consequences).

Many of our pretrained models included more than just the weights so that they could not be loaded anymore with the latest PyTorch version. To fix this problem, we upgraded all our pretrained models to only include the weights so that no error is raised while using them in our framework. If you already downloaded the models, it is recommended to the delete (usually stored in torch.hub.get_dir()) and re-download them.

A nice side effect is that the new models are considerably smaller than the old ones (~ 600 MiB instead of ~2.3 GiB)

v0.0.18: Many tests as examples

We added tests from our internal repository which adds many more examples for the usage of our functions. Please note that many tests require access to internal data and cannot be run directly. They are also not automatically run by GitHub, however, they are always run in our internal GitLab infrastructure so they work in principle.

This release is also compatible with the latest changes of the lazy-imports package.

v0.0.17: xeno-learning

• (New project) xeno-learning: knowledge transfer across species. This includes several new pretrained models for the pig, rat and human species
• Support for Python 3.12
• 3.10 is the new minimal supported Python version
• Switch from flake8/black to ruff
• Support for PyTorch 2.5
• Many style changes and bug fixes from our internal repository

v0.0.16: Housekeeping

This release is compatible with the PyTorch 2.4 release.

v0.0.15: Housekeeping

This release is compatible with the PyTorch 2.2 release.

v0.0.14: Bugfixing

This release includes (besides bug fixes and dependency upgrades) a new SinglePredictor class to create predictions for a trained model in the main process without a consumer/producer architecture

Full Changelog: v0.0.13...v0.0.14

v0.0.13: PyTorch Conference poster release

This release features our PyTorch Conference poster Dealing with I/O bottlenecks in high-throughput model training including the code for the experiments.

Additionally, the code is now compatible with PyTorch 2.1.

v0.0.12: Context paper release

This release features the code and new models for our MICCAI2023 paper Semantic segmentation of surgical hyperspectral images under geometric domain shifts. See also the section in the README for more details.

v0.0.11: HeiPorSPECTRAL dataset release

Latest code to work with the recently released HeiPorSPECTRAL dataset.

Includes many internal updates, e.g PyTorch 2.0 is now supported.