def test_random_transform_generator():
"""
Test random_transform_generator by confirming that it generates
appropriate solutions and output sizes for seeded examples.
"""
# Check shapes are correct Batch Size = 1 - Pass
batch_size = 1
transforms = layer_util.random_transform_generator(batch_size, 0)
assert transforms.shape == (batch_size, 4, 3)
# Check numerical outputs are correct for a given seed - Pass
batch_size = 1
scale = 0.1
seed = 0
expected = tf.constant(
np.array(
[[
[9.4661278e-01, -3.8267835e-03, 3.6934228e-03],
[5.5613145e-03, 9.8034811e-01, -1.8044969e-02],
[1.9651605e-04, 1.4576728e-02, 9.6243286e-01],
[-2.5107686e-03, 1.9579126e-02, -1.2195010e-02],
]],
dtype=np.float32,
)) # shape = (1, 4, 3)
got = layer_util.random_transform_generator(batch_size=batch_size,
scale=scale,
seed=seed)
assert check_equal(got, expected)
def _gen_transforms(self):
"""
Function that generates a random 3D transformation parameters
for a batch of data.
:return: shape = (batch, 4, 3)
"""
return layer_util.random_transform_generator(
batch_size=self._batch_size, scale=self._scale)
## load image
if not os.path.exists(DATA_PATH):
raise ("Download the data using demo_data.py script")
if not os.path.exists(FILE_PATH):
raise ("Download the data using demo_data.py script")
fid = h5py.File(FILE_PATH, "r")
fixed_image = tf.cast(tf.expand_dims(fid["image"], axis=0), dtype=tf.float32)
fixed_image = (fixed_image - tf.reduce_min(fixed_image)) / (
tf.reduce_max(fixed_image) - tf.reduce_min(fixed_image)
) # normalisation to [0,1]
# generate a radomly-affine-transformed moving image
fixed_image_size = fixed_image.shape
transform_random = layer_util.random_transform_generator(batch_size=1, scale=0.2)
grid_ref = layer_util.get_reference_grid(grid_size=fixed_image_size[1:4])
grid_random = layer_util.warp_grid(grid_ref, transform_random)
moving_image = layer_util.resample(vol=fixed_image, loc=grid_random)
# warp the labels to get ground-truth using the same random affine, for validation
fixed_labels = tf.cast(tf.expand_dims(fid["label"], axis=0), dtype=tf.float32)
moving_labels = tf.stack(
[
layer_util.resample(vol=fixed_labels[..., idx], loc=grid_random)
for idx in range(fixed_labels.shape[4])
],
axis=4,
)
## optimisation
def _gen_transforms(self):
return layer_util.random_transform_generator(
batch_size=self._batch_size, scale=self._scale
)