patch_extractor#

deepinv.utils.patch_extractor(imgs, n_patches, patch_size, duplicates=False, position_inds_linear=None)[source]#

This function takes a B x C x H x W tensor as input and extracts n_patches random patches of size C x patch_size x patch_size from each C x H x W image. Hence, the output is of shape B x n_patches x C x patch_size x patch_size.

It returns a tuple of the extracted patches and the linear indices of the patches in the original image.

Parameters:

  • imgs (torch.Tensor) – Images for cutting out patches. Shape batch size x channels x height x width

  • patch_size (int) – size of the patches. The patches are square, so this is the height and width of the patch.

  • n_patches (int) – number of patches to cut out from each image. If -1, all possible patches are cut out.

  • duplicates (bool) – determines if a patch can appear twice.

  • position_inds_linear (torch.Tensor) – allows it to cut patches with specific indices (required for the EPLL reconstruction). dtype of the tensor should be torch.long.

Returns:

tuple of (patches, linear_indices)

Return type:

tuple[Tensor, Tensor]