def test_random_crop_on_list(patch_size, n_inputs):
'''test tf.data compatible random crop function on a list.
NOTE we currently expect the test to fail as there's no extra handling
of lists/tuples/dicts
'''
tf.random.set_seed(17)
img_shape = tuple(2 * val for val in patch_size)
img = np.arange(np.prod(img_shape)).reshape(img_shape)
cropper = random_crop(patch_size)
input_list = [key * img for key in range(1, n_inputs + 1)]
first_patch = cropper(input_list)
for _ in range(10):
patch = cropper(input_list)
# check if patches have the correct shape
for vals in patch:
vals = vals.numpy()
assert vals.ndim == img.ndim
assert np.all(vals.shape == patch_size)
assert len(patch) == n_inputs
# check if all inputs were cropped in the same location
for (ii, first), (jj, second) in pairwise(enumerate(patch, start=1)):
assert np.all(first.numpy() * jj == second.numpy() * ii)
# make sure we're not drawing the same patch over and over
for ii, vals in enumerate(patch):
assert not np.all(vals.numpy() == first_patch[ii].numpy())
def test_random_crop_on_dict(patch_size, n_inputs):
'''test tf.data compatible random crop function on an input_dict.
'''
tf.random.set_seed(17)
img_shape = tuple(2 * val for val in patch_size)
img = np.arange(np.prod(img_shape)).reshape(img_shape)
cropper = random_crop(patch_size)
input_dict = {key: key * img for key in range(1, n_inputs + 1)}
first_patch = cropper(input_dict)
for _ in range(10):
patch = cropper(input_dict)
# check if patches have the correct shape
for key in input_dict.keys():
vals = patch[key].numpy()
assert vals.ndim == img.ndim
assert np.all(vals.shape == patch_size)
assert len(patch.keys()) == n_inputs
# check if all inputs were cropped in the same location
for first_key, second_key in pairwise(patch):
assert np.all(patch[first_key].numpy() *
second_key == patch[second_key].numpy() * first_key)
# make sure we're not drawing the same patch over and over
for key in patch.keys():
assert not np.all(patch[key].numpy() == first_patch[key].numpy())
def test_mismatching_shapes(shapes):
'''test if shape mismatches in the input raise.
'''
patch_size = (13, 13, 1)
inputs = [np.ones(shape) for shape in shapes]
cropper = random_crop(patch_size)
with pytest.raises(tf.errors.InvalidArgumentError):
cropper(inputs)
def test_random_crop_flexible(shapes, patch_size):
'''test cropping with flexible channels.
'''
inputs = [np.ones(shape) for shape in shapes]
cropper = random_crop(patch_size)
patches = cropper(inputs)
patch_size = np.asarray(patch_size)
is_flexible = patch_size == -1
for ii, patch in enumerate(patches):
for dim in range(len(patch_size)):
assert (patch.shape[dim] == patch_size[dim]
or (patch_size[dim] == -1
and patch.shape[dim] == inputs[ii].shape[dim]))