finetuner.tuner.augmentation module#

finetuner.tuner.augmentation.vision_preprocessor(height=224, width=224, default_channel_axis=- 1, target_channel_axis=0, normalize=False, phase='train')[source]#

Randomly augments a Document with tensor field. The method applies flipping, color jitter, cropping, gaussian blur and random rectangle erase to the given image.

Parameters
  • height (int) – image height.

  • width (int) – image width.

  • default_channel_axis (int) – The color channel of the input image, by default -1, the expected input is H, W, C.

  • target_channel_axis (int) – The color channel of the output image, by default 0, the expected output is C, H, W.

  • normalize (bool) – Normalize uint8 image tensor into a float32 image tensor inplace.

  • phase (str) – phase of experiment, either train or validation. At validation phase, will not apply random transformation.