Skip to content

Augmentation

You can specify the data augmentation pipeline in the augmentation section of the configuration file. The augmentation pipeline is a list of transformations that are probabilistically applied to the images in the dataset. The transformations are applied in the order they are specified in the configuration file.

An example of the most minimal non-empty augmentation pipeline is the following:

augmentation:
    enabled: true
    steps:
      - uses: random_crop
        args:
            crop_size: 224

The enabled field is a boolean that indicates whether data augmentation should be applied to the dataset. If enabled is false, the augmentation pipeline is ignored and the dataset is not augmented.

The steps field is a list of transformations that are applied to the dataset. Each transformation is specified by a dictionary with at least the uses field, which indicates the name of the transformation to use.

A step can also have an args field, which is a dictionary of arguments that will be passed to the transformation.

Note

The uses field is case-insensitive, and all hyphens and underscores are ignored, so RANDOM_CROP, RandomCrop, random-crop and random_crop are all equivalent and will use the same RandomCrop transformation.

There are some other fields that can be specified for each augmentation step:

  • probability (optional float, default: 1.0): the probability that the transformation will be applied to the image. If probability is 0.5, the transformation will be applied to half of the images in the dataset.

  • applies_to (optional list of strings or ints, default: all): the list which contains the indices of the images that the transformation will be applied to, relative to the dataset element (which can have more than one image). If applies_to is all, the transformation will be applied to all images in the dataset element. If applies_to is [0, 2], the transformation will be applied to the first and third images in the dataset element. There are some special values for applies_to:

    • all: the transformation will be applied to all images in the dataset element (this is the default value); if this value is used, no other value can be specified.
    • probe: the transformation will be applied to the first image in the dataset element, which by convention is the probe image.
    • live: the transformation will be applied to the second image in the dataset element, which by convention is the live image.

Note

The probability of applying a transformation to an image is on the whole step, not on each image in the dataset element.

For instance, if the dataset element has 3 images, and the step has probability: 0.5 and applies_to: [0, 2], the transformation will be atomically applied to both images with probability 0.5: either both images will be transformed, or none of them will.