config_transforms

transforms

Specifies the transforms to be used during training and during evaluation. ToTensor() and Normalize do not need to be specified, as they are added by default.

Default yaml:

transforms:
  train:
    Resize:
      size: 256
    RandomResizedCrop:
      scale: 0.16 1
      ratio: 0.75 1.33
      size: 227
    RandomHorizontalFlip:
      p: 0.5

  eval:
    Resize:
      size: 256
    CenterCrop:
      size: 227

Example command line modification:

# Use RandomVerticalFlip instead of RandomHorizontalFlip
--transforms {train~SWAP~1: {RandomHorizontalFlip: RandomVerticalFlip}}