def testIdentityGrads(self):
"""Tests that Gradients for 1.0 scale should be ones for some kernels."""
in_shape = [1, 2, 3, 1]
out_shape = [1, 4, 6, 1]
x = np.arange(0, 6).reshape(in_shape).astype(np.float32)
kernel_types = [
'lanczos1', 'lanczos3', 'lanczos5', 'triangle', 'keyscubic'
]
scale = (1.0, 1.0)
translation = (0.0, 0.0)
antialias = True
for kernel_type in kernel_types:
with self.cached_session():
input_tensor = constant_op.constant(x, shape=in_shape)
with backprop.GradientTape() as tape:
tape.watch(input_tensor)
scale_and_translate_out = image_ops.scale_and_translate(
input_tensor,
out_shape[1:3],
scale=constant_op.constant(scale),
translation=constant_op.constant(translation),
kernel_type=kernel_type,
antialias=antialias)
grad = tape.gradient(scale_and_translate_out, input_tensor)[0]
grad_v = self.evaluate(grad)
self.assertAllClose(np.ones_like(grad_v), grad_v)
def testGrads(self):
in_shape = [1, 2, 3, 1]
out_shape = [1, 4, 6, 1]
x = np.arange(0, 6).reshape(in_shape).astype(np.float32)
kernel_types = [
'lanczos1', 'lanczos3', 'lanczos5', 'gaussian', 'box', 'triangle',
'keyscubic', 'mitchellcubic'
]
scales = [(1.0, 1.0), (0.37, 0.47), (2.1, 2.1)]
translations = [(0.0, 0.0), (3.14, 1.19), (2.1, 3.1), (100.0, 200.0)]
for scale in scales:
for translation in translations:
for kernel_type in kernel_types:
for antialias in [True, False]:
with self.cached_session():
input_tensor = constant_op.constant(x,
shape=in_shape)
scale_and_translate_out = image_ops.scale_and_translate(
input_tensor,
out_shape[1:3],
scale=constant_op.constant(scale),
translation=constant_op.constant(translation),
kernel_type=kernel_type,
antialias=antialias)
err = gradient_checker.compute_gradient_error(
input_tensor,
in_shape,
scale_and_translate_out,
out_shape,
x_init_value=x)
self.assertLess(err, 1e-3)
def testIdentityGrads(self):
"""Tests that Gradients for 1.0 scale should be ones for some kernels."""
in_shape = [1, 2, 3, 1]
out_shape = [1, 4, 6, 1]
x = np.arange(0, 6).reshape(in_shape).astype(np.float32)
kernel_types = ['lanczos1', 'lanczos3', 'lanczos5', 'triangle', 'keyscubic']
scale = (1.0, 1.0)
translation = (0.0, 0.0)
antialias = True
for kernel_type in kernel_types:
with self.cached_session():
input_tensor = constant_op.constant(x, shape=in_shape)
with backprop.GradientTape() as tape:
tape.watch(input_tensor)
scale_and_translate_out = image_ops.scale_and_translate(
input_tensor,
out_shape[1:3],
scale=constant_op.constant(scale),
translation=constant_op.constant(translation),
kernel_type=kernel_type,
antialias=antialias)
grad = tape.gradient(scale_and_translate_out, input_tensor)[0]
grad_v = self.evaluate(grad)
self.assertAllClose(np.ones_like(grad_v), grad_v)
def testGrads(self):
in_shape = [1, 2, 3, 1]
out_shape = [1, 4, 6, 1]
x = np.arange(0, 6).reshape(in_shape).astype(np.float32)
kernel_types = [
'lanczos1', 'lanczos3', 'lanczos5', 'gaussian', 'box', 'triangle',
'keyscubic', 'mitchellcubic'
]
scales = [(1.0, 1.0), (0.37, 0.47), (2.1, 2.1)]
translations = [(0.0, 0.0), (3.14, 1.19), (2.1, 3.1), (100.0, 200.0)]
for scale in scales:
for translation in translations:
for kernel_type in kernel_types:
for antialias in [True, False]:
with self.cached_session():
input_tensor = constant_op.constant(x, shape=in_shape)
scale_and_translate_out = image_ops.scale_and_translate(
input_tensor,
out_shape[1:3],
scale=constant_op.constant(scale),
translation=constant_op.constant(translation),
kernel_type=kernel_type,
antialias=antialias)
err = gradient_checker.compute_gradient_error(
input_tensor,
in_shape,
scale_and_translate_out,
out_shape,
x_init_value=x)
self.assertLess(err, 1e-3)
def scale_trans(input_tensor,
scale=scale,
translation=translation,
kernel_type=kernel_type,
antialias=antialias):
# pylint: disable=cell-var-from-loop
return image_ops.scale_and_translate(
input_tensor,
out_shape[1:3],
scale=constant_op.constant(scale),
translation=constant_op.constant(translation),
kernel_type=kernel_type,
antialias=antialias)
