fovi.utils.std_transforms

fovi.utils.std_transforms.get_std_transforms(where, flip, color_jitter, gray, blur, device, dtype, pointcloud_mode=False, solarize=False, normalize=True)[source]

Build standard augmentation pipelines for training.

Creates three transform pipelines (loader, pre-warp, post-warp) that can be applied at different stages of the data processing pipeline.

Parameters:
  • where (str) – Where to apply transforms. One of ‘loader’, ‘pre_warp’, or ‘post_warp’. Determines which pipeline contains the augmentations.

  • flip (bool) – Whether to include random horizontal flip.

  • color_jitter (bool) – Whether to include color jitter augmentation.

  • gray (bool) – Whether to include random grayscale.

  • blur (bool) – Whether to include Gaussian blur.

  • device – Device to place transforms on.

  • dtype – Data type for tensors.

  • pointcloud_mode (bool, optional) – Whether to use pointcloud-compatible transforms. Defaults to False.

  • solarize (bool, optional) – Whether to include random solarization. Defaults to False.

  • normalize (bool, optional) – Whether to include ImageNet normalization. Defaults to True.

Returns:

(loader_transforms, pre_transforms, post_transforms) where each

is either a Compose object or None if empty.

Return type:

tuple