Slideflow

- Fixes test.py, now respects logging level
- Fixes stats.predict_from_layer bug
- Fixes bug in filter_blank where patients with outcome of 0 were being excluded when using linear outcomes
- Fixes errors if slide ID is integer (annotations files now read in str format into pandas DataFrame, rather than auto-detecting dtype)
- Fixes bug where training augmentation strategy set in ModelParams was incorrectly ignored in PyTorch backend (previously forced to always True)
- Fixes 136
- Fixes 134
- Fixes 132
- Addresses 133
- Sets num_parallel_reads=1 and num_workers=1 when calculating activations for DatasetFeatures to ensure records from TFRecord files are correctly ordered
- Changes default augmentation to False when calling sf.io.interleave() or sf.Dataset.torch() or tensorflow()
- Condensed dockerfiles
- Updated DOI (v1.0.4)

- Automatic num_threads for tile extraction based on CPU core count
- Fixes heatmap() batch_size argument (previously ignored)
- Fixes dataset error handling for missing tfrecords
- Fixed Dataset.tfrecord_report()
- Removes previously deprecated vmtouch buffer
- Fixes TMA tile extraction
- Updated QuPath ROI instructions (legacy compatibility)
- Removed Dataset.slide_report() (just do extract_tiles(... dry_run=True)
- Updated sample_actions.py
- Reports split ID when loading a validation split
- TestConfigurator -> TestConfig
- slideflow.statistics -> slideflow.stats

_Doc updates:_
- New "Installation" section
- "Switching backends" section
- Updated "Troubleshooting" section
- Updated wording for what differentiates tutorials 1 & 2
- New tutorial 3 - evaluation & heatmaps
- Doc source cleanup - removed unused images
- Various minor updates
- Collapsed navigation for docs
- Updated github links
- Purple color update
- removed example from bash
- code example improvements

- Fixes bug where model was not saved if val_strategy is None (PyTorch backend)
- Fixes erroneous divide by zero warning with np.log when calculating blur burden
- Fixes bug with Features and real-time normalization (PyTorch)
- Fixes FileNotFoundError when saving a mosaic to the current working directory

- Fixes incorrect requirements.txt
- Test script improvements, setup.py
- Removes gitpython from install_requires, adds click
- TF/PyTorch import fixes, git fix
- Skips CLAM testing if torch not installed
- Adds ubuntu pixman==0.38 repair
- Adds Tensorflow dockerfile & Torch dockerfile
- Adds more testing CLI options to test.py script
- Moves tfrecord2idx to sf.util so it does not import pytorch
- Builds libvips from source with tensorflow dockerfile
- CLAM submodule fixes
- Neptune logging fixes
- Fixes pytorch early stopping bug
- numpy must be <1.21

- Typos in documentation
- Disabled threading in torch chunk loader (functionality still present, not used)
- Critical bug fix in preserved site cross-validation
- New experimental adversarial trainer (PyTorch only; experimental)
- Logging level bug fix
- Multithreading in build_index
- Bug fix with sf.Project creation if dataset configuration not yet set up
- Bug fixes for slides already extracted
- Updated requirements.txt
- Significant improvement in tile extraction when downsampling enabled
- Bug fix in PyTorch loader if a tfrecord has a single tile
- SF_LOGGING_LEVEL utilization for multiprocessing
- Binomial Z now reported in debugging logs for categorical outcomes
- Bug fix for onehot encoded labels (GAN compatibility)

**API Changes**:
- improved neptune logging
- k-fold-preserved-site now requires an annotations header column for determining site, rather than reading characters 5:7 from the patient name. This can be passed via `val_k_fold_header`, which defaults to `'site'`
- default img_format changed to `jpg`
- changed submitter_id -> patient as patient-level identifier in annotations files
- SFP -> P in actions.py
- train(hyperparameters=) -> train(params=)
- hyperparameters.json -> params.json

**Other improvements**:
- Faster tfrecord verification
- Normalizer strategy now logged as a hyperparameter
- Neptune API & workspace read automatically from environmental variables `NEPTUNE_API_KEY` and `NEPTUNE_WORKSPACE`
- Easier custom model training by passing the model class or function directly to `ModelParams.model`
- Adds TIFF EXIF reading improvements for reading MPP from certain files
- Blur detection quality control accessible via WSI.qc() (or by passing `qc='blur'` to `extract_tiles()`)
- Removed duplicate evaluation & predictions
- PyTorch multi-GPU support
- Numpy re-seeding each epoch to improve augmentation randomness
- Updated documentation

**Known issues:**
- Tile extraction may hang if num_workers > 1