def _conv_laplace_2d(tensor):
kernel = np.array([[0., 1., 0.], [1., -4., 1.], [0., 1., 0.]], dtype=np.float32)
kernel = kernel.reshape((3, 3, 1, 1))
if tensor.shape[-1] == 1:
return math.conv(tensor, kernel, padding='VALID')
else:
return math.concat([math.conv(tensor[..., i:i + 1], kernel, padding='VALID') for i in range(tensor.shape[-1])], -1)
def _conv_laplace_3d(tensor):
"""
3D/Cube laplace stencil in 3D+2D [3,3,3,1,1]
array([[[[[ 0.]], [[ 0.]], [[ 0.]]],
[[[ 0.]], [[ 1.]], [[ 0.]]],
[[[ 0.]], [[ 0.]], [[ 0.]]]],
[[[[ 0.]], [[ 1.]], [[ 0.]]],
[[[ 1.]], [[-6.]], [[ 1.]]],
[[[ 0.]], [[ 1.]], [[ 0.]]]],
[[[[ 0.]], [[ 0.]], [[ 0.]]],
[[[ 0.]], [[ 1.]], [[ 0.]]],
[[[ 0.]], [[ 0.]], [[ 0.]]]]]
returns ...
padding explicitly done in laplace(), hence here not needed
"""
kernel = math.to_float([[[0., 0., 0.], [0., 1., 0.], [0., 0., 0.]],
[[0., 1., 0.], [1., -6., 1.], [0., 1., 0.]],
[[0., 0., 0.], [0., 1., 0.], [0., 0., 0.]]])
kernel = kernel.reshape((3, 3, 3, 1, 1))
if tensor.shape[-1] == 1:
return math.conv(tensor, kernel, padding='VALID')
else:
return math.concat([
math.conv(tensor[..., i:i + 1], kernel, padding='VALID')
for i in range(tensor.shape[-1])
], -1)
def _gradient_nd(tensor, padding, relative_shifts):
rank = spatial_rank(tensor)
tensor = math.pad(tensor, _get_pad_width(rank, (-relative_shifts[0], relative_shifts[1])), mode=padding)
components = []
for dimension in range(rank):
lower, upper = _dim_shifted(tensor, dimension, relative_shifts, diminish_others=(-relative_shifts[0], relative_shifts[1]))
components.append(upper - lower)
return math.concat(components, axis=-1)