def _test_partial_shape_correctness(self,
input,
rank,
batch_size,
grid,
interpolation,
boundary,
expected_value=None):
resampler = ResamplerOptionalNiftyRegLayer(interpolation=interpolation,
boundary=boundary)
input_default = tf.random_uniform(input.shape)
if batch_size > 0 and rank > 0:
input_placeholder = tf.placeholder_with_default(
input_default, shape=[batch_size] + [None] * (rank + 1))
elif batch_size <= 0 and rank > 0:
input_placeholder = tf.placeholder_with_default(input_default,
shape=[None] *
(rank + 2))
elif batch_size <= 0 and rank <= 0:
input_placeholder = tf.placeholder_with_default(input_default,
shape=None)
out = resampler(input_placeholder, grid)
with self.cached_session() as sess:
out_value = sess.run(out, feed_dict={input_placeholder: input})
if expected_value is not None:
self.assertAllClose(expected_value, out_value)
def _test_correctness(self, input, grid, interpolation, boundary,
expected_value):
resampler = ResamplerOptionalNiftyRegLayer(interpolation=interpolation,
boundary=boundary)
out = resampler(input, grid)
for use_gpu in self._get_devs():
with self.cached_session(use_gpu=use_gpu) as sess:
out_value = sess.run(out)
self.assertAllClose(expected_value, out_value)
def test_gradient_correctness(self):
if not resampler_module.HAS_NIFTYREG_RESAMPLING:
self.skipTest('Using native NiftyNet resampler; skipping test')
return
for inter in ('LINEAR', 'BSPLINE'):
for b in ('ZERO', 'REPLICATE', 'SYMMETRIC'):
for use_gpu in self._get_devs():
inputs = (
(self.get_3d_input1(as_tensor=False),
[[[-5.2, .25, .25], [.25, .95, .25]],
[[.75, .25, .25], [.25, .25, .75]]]),
(self.get_2d_input(as_tensor=False), [[[.25, .25],
[.25, .78]],
[[.62, .25],
[.25, .28]]]),
)
for np_img, np_u in inputs:
with self.session(use_gpu=use_gpu):
np_u = np.array(np_u)
while len(np_u.shape) < len(np_img.shape):
np_u = np.expand_dims(np_u, axis=2)
img = tf.constant(np_img, dtype=tf.float32)
disp = tf.constant(np_u, dtype=tf.float32)
# multimodal needs addressing
if img.shape.as_list()[-1] > 1:
img = tf.reshape(
img[..., 0],
img.shape.as_list()[:-1] + [1])
warped = ResamplerOptionalNiftyRegLayer(
interpolation=inter, boundary=b)
warped = warped(img, disp)
#warped = tf.reduce_sum(warped)
tgrad, refgrad = tft.compute_gradient(
disp, disp.shape, warped, warped.shape)
error = np.power(tgrad - refgrad, 2).sum()
refmag = np.power(refgrad, 2).sum()
self.assertLessEqual(error, 1e-2 * refmag)