- Updated README
- Updates to workflows: RODI and Finbenthic2
- Added random state option to
01_train_test_split.py
- Explicit setting of
wandb.init()
so that it works in multi-GPU environments
- Feature extraction is now possible in
03_predict.py
. Setting--feature_extraction
to "pooled" or "unpooled" returs a pickled file containing the feature outputs of the DNN, before a classification head. - Feature extraction can be also done with pretrained models, if no checkpoint is passed to prediction script.
- Added possibility of returning logits instead of sigmoid probabilities, using
--return_logits 'True'
in prediction script
- Added an argument
--suffix
to CV prediction combination, so that specification between grouped and non-grouped predictions in the same folder can be distinguished. - Added
--out_prefix
to evaluation script.metrics
by default.
- Removed imsize as a parameter to
Dataset
andLitDataModule
. This is passed viaaug_args
. - Added segmentation module to the package. Documentation coming later.
- Added onnx, onnxruntime, biopython, networkx and pycocotools to environment
- Added FinBenthic1 examples
- Loading pretrained weights without resuming a previous run is now possible
- Last model checkpoint is saved
- Learning rate monitoring
- Fixed bug in TTA. Changes to DataModules were also made in
taxonomist.__init__
- Fixed a bug in grouping script where reference dataset was not checked properly
- Running evaluation without bootstrap is now possible
- Re-designed and simplified comparison script