Things that could need investigating (ongoing list)

[cssprad1]

This is an evolving list of configurations and settings that I have been using. I have not done more than verify these work as intended however they could stand to be looked over and see if it is best for our training needs

• Automatic resizing of data (code)

        self.transform_img = \
            T.Compose([
                RandomResizedCropNP(scale=(0.67, 1.),
                                    ratio=(3. / 4., 4. / 3.)),
                T.ToTensor(),
                T.Resize((config.DATA.IMG_SIZE, config.DATA.IMG_SIZE)),
            ])

• Scaled learning rate (code)

    linear_scaled_lr, linear_scaled_min_lr, linear_scaled_warmup_lr = \
        setup_scaled_lr(config)

• Choice of scheduler (example)

  LR_SCHEDULER:
    NAME: 'multistep'
    GAMMA: 0.1

• Best mask ratio (example)

  MASK_PATCH_SIZE: 32
  MASK_RATIO: 0.6

[cssprad1]

Done with #63