Support transforms that are only applied on the training data

[ungi]

Some transforms should be used for both training and validation data, like resize and normalization. But some transforms like noise, crop, rotation, should only be applied to the training data.

[chriscyyeung]

What do you think about something like this in the config file?

transforms:
  general:
    # Basic transforms, do not modify
    - name: "Transposed"
      params:
        keys: ["image", "label"]
        indices: [2, 0, 1]
    - name: "ToTensord"
    - name: "EnsureTyped"
      params:
        keys: ["image", "label"]
        dtype: "float32"

    # Add additional transforms here
    - name: "Resized"
      params:
        keys: ["image", "label"]
        spatial_size: [128, 128]
  train:
    - name: "RandGaussianNoised"
      params:
        keys: ["image"]
        prob: 0.15
        mean: 0.0
        std: 0.1

The general transforms would be applied to both training and validation images while the train transforms are only applied to the training data.

[ungi]

Looks good to me. Thanks!