0.4.0

Major changes

• AutoAlbument core is rewritten. Now AutoAlbument uses PyTorch Lightning to search for the best augmentation policies.
• Some configuration parameters in search.yaml are changed to work with PyTorch Lightning. The new release contains a new command named autoalbument-migrate which automatically converts an old config file to the new format. If you try to call autoalbument-search with the old config file, you will get instructions on how to use autoalbument-migrate.

0.3.1

Minor changes

• AutoAlbument now shows a RuntimeWarning instead of triggering an assertion when the sum of augmentation probabilities exceeds 1.0.

0.3.0

Bug fixes

• CutoutFixedNumberOfHoles and CutoutFixedSize are now wrapped in a straight-through estimator.
• Fixed a bug that prevented passing a list as an argument to augmentations defined in preprocessing.
• TensorBoarad logs are now saved to a file without a leading ..

Minor changes

• Added example configs for SVHN, ImageNet, and Cityscapes.
• The CIFAR-10 example is now using WideResNet 28-10 as a custom model.
• The default config for classification now uses num_sub_policies: 50 instead of num_sub_policies: 20.
• The default config for semantic segmentation now uses num_sub_policies: 25 instead of num_sub_policies: 20.
• The same parameters for Cutout coordinates are applied to each image in a chunk
• Datasets in example configs are renamed from SearchDataset to a custom class name for that dataset.

0.2.0

Minor changes

• AutoAlbument now uses the instantiate function from Hydra to create instances of Classification and Semantic Segmentation models. This change allows using custom models inside AutoAlbument. See the CIFAR10 example that uses a custom ResNet-18 model.
• AutoAlbument now requires timm>=0.3.2 and segmentation-models-pytorch>=0.1.3.
• All arguments to SearchDataset classes in examples are now defined in Hydra configs.
• Fixed typos in code.

0.1.1

Bug fixes

• Fixed a bug that resulted in the Error calling 'dataset.SearchDataset' : Error loading module 'dataset.SearchDataset' exception when the value of the--config-dir argument was a relative path.

0.1.0

• Major internal refactoring to follow best practices for Hydra configs.
• autoalbument-create will now ask for confirmation to overwrite a file if it already exists.
• The default way of configuring PyTorch Dataset in search.yaml is changed from the dataset_file parameter (which required a path to a Python file with the dataset) to the dataset parameter (which requires a Python class located within PYTHONPATH). The previous option of providing a dataset filepath is also supported.

0.0.4

Bug fixes

• Added the drop_last: True parameter to the DataLoader configuration to prevent exceptions if the latest batch is an incomplete batch (Issue #1).

0.0.3

Changes

• setup.py now contains an updated project description.

Bug fixes

• Template files now correctly included in the distribution.

0.0.1

Initial public release.