DiceLoss Sweep

TODOs

Perform Sweep with CrossEntropyLoss
Perform Sweep with DiceLoss and FocalLoss
🔥 Retrain model on Multi objective Loss function (Combining FocalLoss/ CrossEntropy + Dice).
$L(\alpha) = \alpha FL(input, targ) + (1-\alpha) Dice(input, targ)$ : Add the $\alpha$ param to the sweep params
Augmentation: We could gain some more accuracy by using fastai vision augmentation

# fastai computer vision transforms 
tfms = aug_transforms()

# pass them to the dataloader
SegmentationDataLoaders.from_label_func(
	artifact_dir,
	bs=batch_size,
        fnames=fnames,
        label_func=label_func,
        codes=codes,
        item_tfms=Resize(
            (image_shape[0] // resize_factor, image_shape[1] // resize_factor)),
        valid_pct=validation_split,
        seed=seed,
	batch_tfms=tfms,
        )

Results of the Latest Sweep

We filter results that:
finished correctly
are recent (last day)
have high accuracy and dice