def upsample(x, factor: int):
""" Up sample image by a factor of 2 using nearest neighbour.
Example:
a = C.input_variable((3, 32, 32)
b = UpSampling2D(a)
assert b.shape == (3, 64, 64)
Arguments:
x: input image tensor, assumed (channel, row, col)
Returns:
:class:`~cntk.ops.functions.Function`
"""
# old implementation for upsample by factor 2
# xr = C.reshape(x, (x.shape[0], x.shape[1], 1, x.shape[2], 1))
# xx = C.splice(xr, xr, axis=-1) # axis=-1 refers to the last axis
# xy = C.splice(xx, xx, axis=-3) # axis=-3 refers to the middle axis
# r = C.reshape(xy, (x.shape[0], x.shape[1] * 2, x.shape[2] * 2))
xx = C.splice(*[x for __ in range(factor * 2)], axis=0)
r = C.depth_to_space(xx, factor)
return r
def test_depth_to_space(image_shape, num_channels, block_size, device_id,
precision):
dev = cntk_device(device_id)
from cntk.internal import sanitize_dtype_cntk
input_val = np.array(np.reshape(range(num_channels), (num_channels, 1, 1)),
dtype=PRECISION_TO_TYPE[precision])
input_val = np.tile(input_val, (1, ) + image_shape)
img = C.input_variable(
(num_channels, ) + image_shape,
dtype=sanitize_dtype_cntk(PRECISION_TO_TYPE[precision]))
# Result from depth_to_space node.
depth_to_space_op = C.depth_to_space(img, block_size)
output_test = depth_to_space_op.eval({img: input_val})
# Reference result from simulating depth_to_space with other CNTK ops.
h, w = image_shape
reshape_node = C.reshape(img, (block_size, block_size, num_channels //
(block_size**2), h, w))
transpose_node = C.transpose(reshape_node, [2, 3, 0, 4, 1])
depth_to_space_sim_op = C.reshape(
transpose_node,
(num_channels // (block_size**2), h * block_size, w * block_size))
output_ref = depth_to_space_sim_op.eval({img: input_val})
assert np.array_equal(output_test, output_ref)
def test_space_to_depth(image_shape, num_channels, block_size, device_id, precision):
dev = cntk_device(device_id)
from cntk.internal import sanitize_dtype_cntk
input_val = np.random.randint(low=0, high=100, size=(num_channels,) + image_shape).astype(PRECISION_TO_TYPE[precision])
img = C.input_variable((num_channels,) + image_shape, dtype=sanitize_dtype_cntk(PRECISION_TO_TYPE[precision]))
depth_to_space_op = C.depth_to_space(img, block_size)
space_to_depth_op = C.space_to_depth(depth_to_space_op, block_size)
output_val = np.squeeze(space_to_depth_op.eval({ img : input_val }), 0)
assert np.array_equal(output_val, input_val)
def test_DepthToSpace(tmpdir):
num_channels = 9
block_size = 3
image_shape = (4, 5)
input_val = np.array(np.reshape(range(num_channels), (num_channels, 1, 1)), dtype=np.float32)
input_val = np.tile(input_val, (1,) + image_shape)
input_val.shape = (1,) + input_val.shape
img = C.input_variable((num_channels,) + image_shape, dtype=np.float32)
model = C.depth_to_space(img, block_size)
verify_one_input(model, input_val, tmpdir, 'DepthToSpace')
def test_depth_to_space(image_shape, num_channels, block_size, output_ref, device_id, precision):
dev = cntk_device(device_id)
from cntk.internal import sanitize_dtype_cntk
input_val = np.array(np.reshape(range(num_channels), (num_channels, 1, 1)), dtype=PRECISION_TO_TYPE[precision])
input_val = np.tile(input_val, (1,) + image_shape)
img = C.input_variable((num_channels,) + image_shape, dtype=sanitize_dtype_cntk(PRECISION_TO_TYPE[precision]))
depth_to_space_op = C.depth_to_space(img, block_size)
output_test = depth_to_space_op.eval({ img : input_val })
assert np.array_equal(output_test, output_ref)
def test_space_to_depth(image_shape, num_channels, block_size, device_id, precision):
dev = cntk_device(device_id)
from cntk.internal import sanitize_dtype_cntk
input_val = np.random.randint(low=0, high=100, size=(num_channels,) + image_shape).astype(PRECISION_TO_TYPE[precision])
img = C.input_variable((num_channels,) + image_shape, dtype=sanitize_dtype_cntk(PRECISION_TO_TYPE[precision]))
depth_to_space_op = C.depth_to_space(img, block_size)
space_to_depth_op = C.space_to_depth(depth_to_space_op, block_size)
output_val = np.squeeze(space_to_depth_op.eval({ img : input_val }), 0)
assert np.array_equal(output_val, input_val)
def test_DepthToSpace(tmpdir):
num_channels = 9
block_size = 3
image_shape = (4, 5)
input_val = np.array(np.reshape(range(num_channels), (num_channels, 1, 1)), dtype=np.float32)
input_val = np.tile(input_val, (1,) + image_shape)
input_val.shape = (1,) + input_val.shape
img = C.input_variable((num_channels,) + image_shape, dtype=np.float32)
model = C.depth_to_space(img, block_size)
verify_one_input(model, input_val, tmpdir, 'DepthToSpace')
def test_depth_to_space(image_shape, num_channels, block_size, output_ref, device_id, precision):
dev = cntk_device(device_id)
from cntk.internal import sanitize_dtype_cntk
input_val = np.array(np.reshape(range(num_channels), (num_channels, 1, 1)), dtype=PRECISION_TO_TYPE[precision])
input_val = np.tile(input_val, (1,) + image_shape)
img = C.input_variable((num_channels,) + image_shape, dtype=sanitize_dtype_cntk(PRECISION_TO_TYPE[precision]))
depth_to_space_op = C.depth_to_space(img, block_size)
output_test = depth_to_space_op.eval({ img : input_val })
assert np.array_equal(output_test, output_ref)
def test_depth_to_space(image_shape, num_channels, block_size, device_id, precision):
dev = cntk_device(device_id)
from cntk.internal import sanitize_dtype_cntk
input_val = np.array(np.reshape(range(num_channels), (num_channels, 1, 1)), dtype=PRECISION_TO_TYPE[precision])
input_val = np.tile(input_val, (1,) + image_shape)
img = C.input_variable((num_channels,) + image_shape, dtype=sanitize_dtype_cntk(PRECISION_TO_TYPE[precision]))
# Result from depth_to_space node.
depth_to_space_op = C.depth_to_space(img, block_size)
output_test = depth_to_space_op.eval({ img : input_val })
# Reference result from simulating depth_to_space with other CNTK ops.
h, w = image_shape
reshape_node = C.reshape(img, (block_size, block_size, num_channels // (block_size**2), h, w))
transpose_node = C.transpose(reshape_node, [2, 3, 0, 4, 1])
depth_to_space_sim_op = C.reshape(transpose_node, (num_channels // (block_size**2), h * block_size, w * block_size))
output_ref = depth_to_space_sim_op.eval({ img : input_val })
assert np.array_equal(output_test, output_ref)
