- Training can now perform validation checks mid-epoch, at every X batches, defined by the argument "validate_on_batch"
- If HP "early_stop" is True, will stop training if exponential moving average (EMA) of validation accuracy does not increase after 3 successive validation checks (previously checks were performed at epoch end)
- Separated mosaic functions in ActivationsVisualizer into separate mosaic.Mosaic class
- Separated UMAP functions from ActivationsVisualizer into separate statistics.TFRecordUMAP class
- Streamlined Dataset class with apply_filters function
- Added ability to customize model regularizer, including L2 implementation
- Added hyperparameters: hidden_layer_width, trainable_layers, L2_weight
- Added training arguments max_ and min_tiles_per_slide (either filters out TFRecords not meeting criteria or caps maximum of tiles to take)
- Additional arguments for model.train(): ema_smoothing (df=2) and ema_observations (df=8)
- By specifying annotation header_x and header_y, SFP.generate_mosaic_from_annotations will create a mosaic map from prespecified coordinates in the annotations file