def test_random_shapes_and_kernel_sizes(self):
shapes = [(6, 5), (5, 6), (6, 6), (11, 1), (1, 11), (0, 1), (1, 0),
(0, 0)]
kernel_sizes = [1, 2, 3, 5]
nb_channels_lst = [None, 1, 3, 4]
# 8*(4*4)*4 = 512 subtests
gen = itertools.product(shapes, nb_channels_lst)
for shape_nochan, nb_channels in gen:
shape = shape_nochan
if nb_channels is not None:
shape = tuple(list(shape) + [nb_channels])
image = np.zeros(shape, dtype=np.uint8)
for ksize_h, ksize_w in itertools.product(kernel_sizes,
kernel_sizes):
with self.subTest(shape=shape,
ksize_h=ksize_h,
ksize_w=ksize_w):
image_pooled = ia.avg_pool(image, (ksize_h, ksize_w))
shape_expected = image_pooled.shape
shape_observed = iapooling._compute_shape_after_pooling(
shape, ksize_h, ksize_w)
assert shape_observed == shape_expected
def _pool_image(self, image, kernel_size_h, kernel_size_w):
return ia.avg_pool(image, (kernel_size_h, kernel_size_w))
def _pool_image(self, image, kernel_size_h, kernel_size_w):
# TODO extend avg_pool to support pad_mode and set it here
# to reflection padding
return ia.avg_pool(image,
(max(kernel_size_h, 1), max(kernel_size_w, 1)),
cval=128)
def _pool_image(self, image, kernel_size_h, kernel_size_w):
return ia.avg_pool(image,
(max(kernel_size_h, 1), max(kernel_size_w, 1)))
