All notable changes to this project will be documented in this file.
The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.
0.10.1 - 2021-05-26
- Replace outdated import from setup.py
0.10.0 - 2021-05-26
- 3D convolution support.
- Support for pytorch 1.8.1
- Removed support for pytorch versions < 1.8
0.9.1 - 2020-10-27
- Fixed the conda build for pytorch versions < 1.5
0.9.0 - 2020-10-26
- Support for Python 3.8
- Suport for PyTorch 1.5
- Fixed normalization for multi-channel input and output.
- Support for PyTorch 1.1
0.8.0 - 2020-03.10
- Weights access when pruning in pytorch 1.4
- torchvision dependency
- sacred dependency
msd_pytorch.relu_inplace
- command-line interface
- old MSDModule
- stitch functions and modules
- Bug in relabeling code in ImageDataset.
0.7.2 - 2019-07-30
- Support for multi-gpu execution. Use
parallel=True
when constructing aMSDRegressionModel
orMSDSegmentationModel
.
- Make
model.forward()
more memory-efficient.
0.7.1 - 2019-05-27
- Add
weights_path
command-line argument to msd executable to indicate where to store final weights. - Add
MSDBlock2d
: this is a faster and slightly more memory efficient implementation of the same MSD component. Many thanks to Jonas Adler for suggesting this way of structuring the code!
- The MSD models use
MSDBlock2d
implementation by default now.
- Initial release.