Slideflow

- Enables extracting tiles at downsample/magnification level without resizing; use by setting `tile_um` equal to a string of format "[int/float]x", such as "10x", "40x", "2.5x". Slides will log an error if there is no matching downsample level at that layer.
- Vastly reduced time and space required for testing, with switch to tile_px=71, tile_um=1208, and mobilenet_v2 architecture
- Added single linear outcome testing to TestSuite
- TestSuite train testing now ensures results are properly returned from `P.train()`
- P.train() now returns a `dict` instead of a multiprocessing `DictProxy`
- Added unit test framework to TestSuite; tests are automatically run with `test.py`. Individual unit tests located in slideflow.test
- Optimized `Dataset.labels()` to make better use of pandas DataFrame functionality
- Added first unit tests in dataset_test.py, includes tests for Dataset.labels()
- Added type hints throughout
- Adds `SlideMap.filter_index()` method which filters x, y, point_meta, and labels of a SlideMap using the given indices
- Adds `method` argument to `SlideMap.neighbors`, allowing neighbors to be calculated from either UMAP coordinates (`'map'`) or feature vectors (`'features'`)
- Refactored PyTorch training loop within `Trainer`. Functionality remains identical, with cleaner organization and reduced code redundancy.
- Import optimization (delays importing seaborn and matplotlib)
- Added `Dataset.thumbnails()`, deprecated `Project.generate_thumbnails()`
- Added `auto_dataset` function decorator for Project functions; preparation for future removal of `filters` and `filter_blank` arguments in Project functions (instead, pass a filtered `Dataset` object)
- Deprecates` Dataset.split_patients_list()`
- Deprecated `Dataset.training_validation_split()`, use `.train_val_split()`
- Deprecated `Project.generate_tfrecord_heatmap()`. Instead, do `Dataset.tfrecord_heatmap()`
- Removed legacy, unused functions in sf.norm.utils
- Clearer error message when attempting to use a normalizer prior to calling `.fit()`
- Added pool.close() for several ThreadPools
- Fixes `sf.util.multi_warn`
- Fixed typo in Tutorial 2
- Fixed docstring typos in `Project.evaluate()` and `Project.predict()`
- Fixes erroneous error log reporting that training had failed when validation was not performed
- Fixes shape error during metrics calculation
- Fixes false positive MismatchImageFormatsError
- Fixed edge case where Torch TFRecord reading was skipping TFRecords with fewer records than the number of shards/workers
- Fixes rare runtime error when exiting an PyTorch dataloader iteration early
- Fixes Neptune logging error before training starts
- Docstring updates

- Documentation update: Tutorial 5 (mosaic maps)
- Adds `use_norm` argument to `Project.generate_mosaic()`. Controls whether image tiles are normalized prior to display using the training normalizer. If False, shows raw images. Defaults to True (prior behavior)
- New `Dataset.img_format` property which verifies all tfrecords have the same image format (PNG/JPG), returning the image format
- Datasets will now raise `MismatchedImageFormatsError` if tfrecords have mixed image formats
- Model training now saves `img_format` property in `params.json`
- Slide-level prediction (via heatmaps) now extracts tiles using the same image format (PNG/JPG) as the trained model
- Unified `sf.io.detect_tfrecord_format()` between backends; no longer imports torch/tensorflow
- The first argument of `sf.io.detect_tfrecord_format()` now returns a list of features rather than a dict
- Fixes bug where eval and predictions were saved in the model directory, not eval directory
- Fixes python-spams install in Tensorflow dockerfile
- Fixes bug where uncertainty was not properly stored in `DatasetFeatures`
- Fixed documentation typo (on Troubleshooting; for debugging, set environmental variable `SF_LOGGING_LEVEL=10`)

- Updated README
- Documentation update: custom training loops
- Documentation update: uncertainty quantification
- Documentation update: new figures
- Documentation update: tutorial 4 (custom architecture)
- Slide PDF report now shows number of tiles extracted
- Fixes overlapping thumbnails in PDF reports
- Fixes broken numpy dependencies due to conflict between `numba` and `spams`

- Fix bug when building Features from model in TF backend
- Fix matplotlib warning when making 3D plot
- Fix bug where models were not being saved if early stopping was triggered
- Tensorflow 2.9 compatibility update

- Enables extracting tiles at downsample/magnification level without resizing; use by setting tile_um equal to a string of format "[int/float]x", such as "10x", "40x", "2.5x"
- New `Dataset.img_format` property which verifies all tfrecords have the same image format, returning the image format
- Datasets will raise `ValueError` if mismatched image formats are found in tfrecords
- Model training now saves `img_format` property in `params.json`
- Slide-level prediction (via heatmaps) now extracts tiles using the same image format (PNG or JPG) as the trained model
- Speeds up test suite by switching to tile_px=76 and tile_um=1208
- Slides will log an error if there is no matching downsample level at that layer
- Fixes edge case where Torch TFRecord reading was skipping TFRecords with fewer records than the number of shards/workers
- Unified sf.io.detect_tfrecord_format() between backends; no longer imports torch/tensorflow
- The first argument of sf.io.detect_tfrecord_format() now returns a list of features, rather than a dictionary
- Enables `include_top=True|False` with PyTorch backend
- Improves model compatibility with PyTorch backend
- Fixes tile extraction PDF bug (error with displaying log(blur_burden))
- Neptune logging fix

- Fixes bug where only one tile per tfrecord was included in mosaic report
- Fixes bug when loading SlideMap from cache
- Fixes normalized images always saved as JPG during tile extraction, even if img_format = PNG
- Fixes neptune logging during evaluation
- Fixes TypeError when training with 3+ outcomes (due to error in sf.Dataset)
- Fixes pandas dtype error for CLAM submodule when slide names are integers
- Fixes DatasetFeatures bug if categorical outcomes are integers
- Fixes bug where IndexError is raised when using filters with mosaic map, or SlideMap.filter()
- Fixes 131
- Sets progress bars to ncols=80
- Adds error handling if Mosaic is created with no tfrecords
- Improves error handling during slide loading (e.g. missing thumbnails)
- DatasetFeatures.normalizer is now an instance of slideflow.StainNormalizer rather than a string and is instanced from hp.get_normalizer()
- Documentation updates: mosaic / layer activations
- stats.py docstring updates