keras_utils module

Common multi-backend Keras utilities

class lib.keras_utils.ColorSpaceConvert(from_space: str, to_space: str)

Bases: object

Transforms inputs between different color spaces on the GPU

Notes

The following color space transformations are implemented:

  • rgb to lab

  • rgb to xyz

  • srgb to _rgb

  • srgb to ycxcz

  • xyz to ycxcz

  • xyz to lab

  • xyz to rgb

  • ycxcz to rgb

  • ycxcz to xyz

Parameters:

  • from_space (str) – One of “srgb”, “rgb”, “xyz”

  • to_space (str) – One of “lab”, “rgb”, “ycxcz”, “xyz”

Raises:

ValueError – If the requested color space conversion is not defined

lib.keras_utils.frobenius_norm(matrix: Tensor, axis: int = -1, keep_dims: bool = True, epsilon: float = 1e-15) → Tensor

Frobenius normalization for Keras Tensor

Parameters:

  • matrix (Tensor) – The matrix to normalize

  • axis (int, optional) – The axis to normalize. Default: -1

  • keep_dims (bool, Optional) – Whether to retain the original matrix shape or not. Default:True

  • epsilon (flot, optional) – Epsilon to apply to the normalization to preven NaN errors on zero values

Returns:

The normalized output

Return type:

Tensor

lib.keras_utils.replicate_pad(image: Tensor, padding: int) → Tensor

Apply replication padding to an input batch of images. Expects 4D tensor in BHWC format.

Notes

At the time of writing Keras/Tensorflow does not have a native replication padding method. The implementation here is probably not the most efficient, but it is a pure keras method which should work on TF.

Parameters:

  • image (Tensor) – Image tensor to pad

  • pad (int) – The amount of padding to apply to each side of the input image

Returns:

The input image with replication padding applied

Return type:

Tensor