Slideflow

- PyTorch intermediate model layers can now be accessed with dot syntax in `sf.Features` (e.g. `'classifier.0'`)
- In the Tensorflow backend, `sf.Features` interface now supports more intermediate layers, including layers not inside the nested core model (such as hidden layers and prelogits)
- Fix for training Inception models in PyTorch backend (disables auxillary classifier and raises warning)
- Fix for `sf.Features` in PyTorch backend for some models when `include_top=False`
- `Project.generate_features_for_clam()` moves model to GPU when generating predictions in PyTorch backend
- New unit tests for models, which verifies that all architectures are built correctly (with and without `include_top`) and can successfully be loaded into `sf.Features`.
- Fixes bug with Macenko normalizer with Heatmaps
- Fixes for Otsu's thresholding on JPEG slides
- Fixes bug where PyTorch heatmaps were not displayed correctly, as the previously displayed logits were pre-softmax

- Switches progress bars and logging to `rich` (requires `rich` package)
- Removes pixman repair for PyTorch dockerfile (not needed)
- Improved error handling for invalid validation settings provided to `Project.train()`
- Improved error handling in `Mosaic` for images that can't be normalized
- Add error checking for mismatched `img_format` for `Project.evaluate()` and `.predict()`
- Safer filename handling
- Documentation & docstring updates for GAN functions
- `SlideMap` support for `DatasetFeatures` lacking logits
- More verbose output when making heatmaps
- Fix `SlideMap` figures missing legends
- Fix for dealing with empty `results_log.csv`
- Fix Tensorflow-native Reinhard mask normalizer
- Fix for eigen decomposition errors for Macenko normalizer (Tensorflow backend)
- Fix for EfficientNet models (Tensorflow)
- Fix for handling `ExtractionReport` when tile extraction fails
- Fix deprecation warning for `tf.data.experimental.sample_from_datasets()`
- Fix for predicting from multi-outcome models
- Update StyleGAN2 submodule commit
- Fix for edge case handling fit.auroc == None or np.nan
- Pin requirement `protobuf<=3.20`

- Optimizations for stain normalizers as described above. **The Vahadane and Macenko implementations changed with optimization, and may yield slightly different results than previously.**
- Updated documentation & expanded stain normalization section
- Fixed Dockerfiles and broken pixman installation

- SlideReports now include grayspace/whitespace fraction for each extracted image, in the DataFrame `SlideReport.locations`
- `include_loc` argument removed from `.extract_tiles()` functions, as location information is now always included in the returned dict
- Performance improvements in ROI calculations and slide verification
- Previously, `.extract_tiles()` returned an `ExtractionReport` summarizing what tiles were extracted. Thus report is saved as a PDF in the tile extraction directory. The returned report, however, only summarized tile extraction of the last dataset source. If multiple dataset sources had tiles extracted, on the last source was reflected in this report. Now, for simplicity, this function returns a Dict mapping slide paths to `SlideReport` for each slide, including slides from all dataset sources.
- Enable creating `Dataset` from config dictionary rather than path
- Fix for `dataset.labels()` when slide column has empty values in annotations
- `sf.stats.metrics.ClassifierMetrics` now autofits
- Fix for using saliency maps with PyTorch
- Fix for normalizer fit not saving to `params.json` in PyTorch backend
- Bug fix for `Features()` in PyTorch backend
- Removes Tensorflow requirement from GAN interpolation function
- PyTorch 1.12 compatibility update
- Adds ninja as dependency (required for StyleGAN2)
- Documentation & docstring updates

- Fixes broken submodule which failed to be included with the 1.2.0 wheel

- Log a warning if a `Dataset` cannot find a given configuration source. Previously, missing sources would be quietly ignored.
- Allow using Slideflow without libvips/pyvips installed.
- Fix bug where models were not being saved if early stopping was triggered.
- Compatibility update for Tensorflow 2.9
- Fix logging when testing multiple/single linear outcomes during testing
- Fix logging message when generating features
- Fix type hints
- Fix bug when building Features from model in TF backend
- Fix training bug when `validate_on_batch` is set to 0 in the pytorch backend
- Fix matplotlib warning when making 3D plot
- Fix bug where logging level was fixed to WARNING after running unit tests