3

I am using automl from google using a custom dataset. The dataset consists of images collected by me. However, manually labelling the image take some time, so I would like to enlarge the dataset by image augmentation, such as rotation and blurring. Does automl perform augmentation behind the screen automatically?

Brian Chiu
  • 33
  • 3

1 Answers1

3

AutoML does few types of data augmentation. This is implementation detail and may change in the future without notice, basic augmentations that are used are:

  • random resizes / crops
  • random flip left right
  • random color and brightness distortions
  • more may be used / added in the future

If doing data augmentation on your side please follow best practices:

  • if you augment image - please put all augmentations of the same image in the same part of the dataset (TRAIN, VALIDATION, TEST) - otherwise the model can overfit without noticing (if almost the same image is in TRAIN and VALIDATION set)
  • do transformations that are meaningful in your context - for example rotations - if in typical use-case you don't get rotated images or objects that are rotated - then training model to detect rotated images may not help your application (for example if you never see people who are upside down in your real application - than training with people who are upside down may not benefit your end model).
Michal K
  • 205
  • 1
  • 5
  • 2
    Can you please share the link to this implementation or documentation? How can one track the augmentation approaches they do behind the scenes and across versions – Qaisar Rajput Sep 19 '19 at 15:11