def testCompareGpuVsCpu(self):
in_shape = [1, 4, 6, 3]
out_shape = [1, 8, 16, 3]
for nptype in self.TYPES:
x = np.arange(0,
np.prod(in_shape)).reshape(in_shape).astype(nptype)
for align_corners in [True, False]:
with self.test_session(use_gpu=False):
input_tensor = constant_op.constant(x, shape=in_shape)
resize_out = image_ops.resize_nearest_neighbor(
input_tensor,
out_shape[1:3],
align_corners=align_corners)
grad_cpu = gradient_checker.compute_gradient(
input_tensor,
in_shape,
resize_out,
out_shape,
x_init_value=x)
with self.test_session(use_gpu=True):
input_tensor = constant_op.constant(x, shape=in_shape)
resize_out = image_ops.resize_nearest_neighbor(
input_tensor,
out_shape[1:3],
align_corners=align_corners)
grad_gpu = gradient_checker.compute_gradient(
input_tensor,
in_shape,
resize_out,
out_shape,
x_init_value=x)
self.assertAllClose(grad_cpu, grad_gpu, rtol=1e-5, atol=1e-5)
def resize_image(image,
height_factor,
width_factor,
data_format='NHWC',
interpolation='nearest'):
if data_format[-1] == 'C':
row, col = 1, 2
else:
row, col = 2, 3
old_shape = int_shape(image)
spatial = array_ops.shape(image)[row:col + 1]
spatial *= array_ops.constant(np.array([height_factor, width_factor]),
'int32')
if data_format[-1] != 'C':
image = transpose_to_channels_last(image)
if interpolation == 'nearest':
image = image_ops.resize_nearest_neighbor(image, size=spatial)
elif interpolation == 'bilinear':
image = image_ops.resize_bilinear(image, size=spatial)
else:
raise ValueError("Unknown interpolation named " + interpolation)
height = old_shape[row] or old_shape[row] * height_factor
width = old_shape[col] or old_shape[col] * width_factor
if data_format[-1] != 'C':
image = transpose_to_channels_first(image)
image.set_shape(old_shape[:2] + (height, width))
else:
image.set_shape((old_shape[0], ) + (height, width) + (old_shape[-1], ))
return image
def testCompareGpuVsCpu(self):
in_shape = [1, 4, 6, 3]
out_shape = [1, 8, 16, 3]
for nptype in self.TYPES:
x = np.arange(0, np.prod(in_shape)).reshape(in_shape).astype(nptype)
for align_corners in [True, False]:
with self.test_session(use_gpu=False):
input_tensor = constant_op.constant(x, shape=in_shape)
resize_out = image_ops.resize_nearest_neighbor(
input_tensor, out_shape[1:3], align_corners=align_corners)
grad_cpu = gradient_checker.compute_gradient(
input_tensor, in_shape, resize_out, out_shape, x_init_value=x)
with self.test_session(use_gpu=True):
input_tensor = constant_op.constant(x, shape=in_shape)
resize_out = image_ops.resize_nearest_neighbor(
input_tensor, out_shape[1:3], align_corners=align_corners)
grad_gpu = gradient_checker.compute_gradient(
input_tensor, in_shape, resize_out, out_shape, x_init_value=x)
self.assertAllClose(grad_cpu, grad_gpu, rtol=1e-5, atol=1e-5)
def testShapeIsCorrectAfterOp(self):
in_shape = [1, 2, 2, 1]
out_shape = [1, 4, 6, 1]
for nptype in self.TYPES:
x = np.arange(0, 4).reshape(in_shape).astype(nptype)
input_tensor = constant_op.constant(x, shape=in_shape)
resize_out = image_ops.resize_nearest_neighbor(
input_tensor, out_shape[1:3])
with self.cached_session(use_gpu=True):
self.assertEqual(out_shape, list(resize_out.get_shape()))
resize_out = self.evaluate(resize_out)
self.assertEqual(out_shape, list(resize_out.shape))
def testGradFromResizeToSmallerInBothDims(self):
in_shape = [1, 4, 6, 1]
out_shape = [1, 2, 3, 1]
for nptype in self.TYPES:
x = np.arange(0, 24).reshape(in_shape).astype(nptype)
with self.test_session(use_gpu=True):
input_tensor = constant_op.constant(x, shape=in_shape)
resize_out = image_ops.resize_nearest_neighbor(input_tensor,
out_shape[1:3])
err = gradient_checker.compute_gradient_error(
input_tensor, in_shape, resize_out, out_shape, x_init_value=x)
self.assertLess(err, 1e-3)
def testGradFromResizeToSmallerInBothDims(self):
in_shape = [1, 4, 6, 1]
out_shape = [1, 2, 3, 1]
for nptype in self.TYPES:
x = np.arange(0, 24).reshape(in_shape).astype(nptype)
with self.cached_session(use_gpu=True):
input_tensor = constant_op.constant(x, shape=in_shape)
resize_out = image_ops.resize_nearest_neighbor(input_tensor,
out_shape[1:3])
err = gradient_checker.compute_gradient_error(
input_tensor, in_shape, resize_out, out_shape, x_init_value=x)
self.assertLess(err, 1e-3)
def testShapeIsCorrectAfterOp(self):
in_shape = [1, 2, 2, 1]
out_shape = [1, 4, 6, 1]
for nptype in self.TYPES:
x = np.arange(0, 4).reshape(in_shape).astype(nptype)
with self.test_session(use_gpu=True) as sess:
input_tensor = constant_op.constant(x, shape=in_shape)
resize_out = image_ops.resize_nearest_neighbor(input_tensor,
out_shape[1:3])
self.assertEqual(out_shape, list(resize_out.get_shape()))
resize_out = sess.run(resize_out)
self.assertEqual(out_shape, list(resize_out.shape))
def testExceptionThrowing(self, align_corners, half_pixel_centers,
data_type):
with self.session(), test_util.force_gpu():
input_image = array_ops.zeros((1, 2, 2, 1), dtype=data_type)
with backprop.GradientTape() as tape:
tape.watch(input_image)
output_image = image_ops.resize_nearest_neighbor(
input_image, (3, 3),
align_corners=align_corners,
half_pixel_centers=half_pixel_centers)
with self.assertRaisesRegex(
errors.UnimplementedError,
'A deterministic GPU implementation of ResizeNearestNeighborGrad'
+ ' is not currently available.'):
gradient = tape.gradient(output_image, input_image)
self.evaluate(gradient)
def resize_nn(t, shape=out_shape, align_corners=align_corners):
return image_ops.resize_nearest_neighbor(
t, shape[1:3], align_corners=align_corners)
def resize_nn(t, shape=out_shape):
return image_ops.resize_nearest_neighbor(t, shape[1:3])
def call(self, inputs, training=None):
# return tf.image.resize(inputs, self.size)
# return resize_images(inputs, self.size)
return image_ops.resize_nearest_neighbor(inputs, self.size)
