niftynet.layer.layer_util module

check_spatial_dims(input_tensor, criteria)[source]

valid each of the spatial dims against criteria criteria can be a lambda function e.g. lambda x : x > 10 checks whether each dim is greater than 10


e.g. given an input tensor [Batch, X, Y, Z, Feature] the spatial rank is 3


This function generates a trivial kernel with all 0s except for the element in its spatial center e.g. trivial_kernel((3, 3, 1, 1,)) returns a kernel of:

[[[[0]], [[0]], [[0]]],
 [[[0]], [[1]], [[0]]],
 [[[0]], [[0]], [[0]]]]

kernel_shape[-1] and kernel_shape[-2] should be 1, so that it operates on the spatial dims only. However, there is no exact spatial centre if np.any((kernel_shape % 2) == 0). This is fine in many cases as np.sum(trivial_kernel(kernel_shape)) == 1

expand_spatial_params(input_param, spatial_rank, param_type=<type 'int'>)[source]

Expand input parameter e.g., kernel_size=3 is converted to kernel_size=[3, 3, 3] for 3D images (when spatial_rank == 3).

check_divisible_channels(input_tensor, n_channel_splits)[source]

Check if the number of channels (last dim) of the input tensor is divisible by n_channel_splits. If True, returns n_input_channels / n_channel_splits, raises AssertionError otherwise

  • input_tensor
  • n_channel_splits

n_input_channels / n_channel_splits