def resnet_down_block(u, ul, us, uls):
us = us.copy()
uls = uls.copy()
for _ in range(nr_resnet):
u = ResnetDown()(u, us.pop())
ul = ResnetDownRight()(ul, np.concatenate((u, uls.pop()), -1))
return u, ul, us, uls
def up_pass(images):
images = np.concatenate((images, np.ones(images.shape[:-1] + (1, ))),
-1)
us = [down_shift(ConvDown(filter_shape=(2, 3))(images))]
uls = [
down_shift(ConvDown(filter_shape=(1, 3))(images)) +
right_shift(ConvDownRight(filter_shape=(2, 1))(images))
]
us, uls = ResnetUpBlock()(us, uls)
us.append(HalveDown()(us[-1]))
uls.append(HalveDownRight()(uls[-1]))
us, uls = ResnetUpBlock()(us, uls)
us.append(HalveDown()(us[-1]))
uls.append(HalveDownRight()(uls[-1]))
return ResnetUpBlock()(us, uls)
def right_shift(input):
h, _, c = input.shape
return np.concatenate((np.zeros((h, 1, c)), input[:, :-1]), 1)
def down_shift(input):
_, w, c = input.shape
return np.concatenate((np.zeros((1, w, c)), input[:-1]), 0)
def concat_elu(x, axis=-1):
return elu(np.concatenate((x, -x), axis))