def _assertBackwardOpMatchesExpected(self, input_np, warp_np,
grad_output_np, expected_grad_data,
expected_grad_warp):
with self.session() as sess, self.test_scope():
input_image = array_ops.placeholder(input_np.dtype)
warp = array_ops.placeholder(warp_np.dtype)
grad_output = array_ops.placeholder(grad_output_np.dtype)
grad_data, grad_warp = gen_resampler_ops.resampler_grad(
input_image, warp, grad_output)
grad_data_tf, grad_warp_tf = sess.run([grad_data, grad_warp], {
input_image: input_np,
warp: warp_np,
grad_output: grad_output_np
})
self.assertAllCloseAccordingToType(expected_grad_warp,
grad_warp_tf,
half_rtol=1e-2,
bfloat16_rtol=3e-2)
self.assertAllCloseAccordingToType(expected_grad_data,
grad_data_tf,
half_rtol=1e-2,
bfloat16_rtol=3e-2)
def _assertBackwardOpMatchesExpected(self, input_np, warp_np, grad_output_np,
expected_grad_data, expected_grad_warp):
with self.cached_session() as sess, self.test_scope():
input_image = array_ops.placeholder(input_np.dtype)
warp = array_ops.placeholder(warp_np.dtype)
grad_output = array_ops.placeholder(grad_output_np.dtype)
grad_data, grad_warp = gen_resampler_ops.resampler_grad(
input_image, warp, grad_output)
grad_data_tf, grad_warp_tf = sess.run([grad_data, grad_warp], {
input_image: input_np,
warp: warp_np,
grad_output: grad_output_np
})
self.assertAllCloseAccordingToType(
expected_grad_warp, grad_warp_tf, half_rtol=1e-2, bfloat16_rtol=3e-2)
self.assertAllCloseAccordingToType(
expected_grad_data, grad_data_tf, half_rtol=1e-2, bfloat16_rtol=3e-2)
def _resampler_grad(op, grad_output): data, warp = op.inputs grad_output_tensor = ops.convert_to_tensor(grad_output, name="grad_output") return gen_resampler_ops.resampler_grad(data, warp, grad_output_tensor)
def _resampler_grad(op, grad_output):
data, warp = op.inputs
grad_output_tensor = ops.convert_to_tensor(grad_output, name="grad_output")
return gen_resampler_ops.resampler_grad(data, warp, grad_output_tensor)