def _test_op_backward_pass(self, on_gpu, dtype, tol):
np.random.seed(13)
data_width = 5
data_height = 4
data_channels = 3
warp_width = 2
warp_height = 6
batch_size = 3
warp = _make_warp(batch_size, warp_height, warp_width,
dtype.as_numpy_dtype)
data_shape = (batch_size, data_height, data_width, data_channels)
data = np.random.rand(*data_shape).astype(dtype.as_numpy_dtype)
with self.test_session(use_gpu=on_gpu, force_gpu=False):
data_tensor = constant_op.constant(data)
warp_tensor = constant_op.constant(warp)
output_tensor = resampler.resampler(data=data_tensor,
warp=warp_tensor)
grads = test.compute_gradient([data_tensor, warp_tensor], [
data_tensor.get_shape().as_list(),
warp_tensor.get_shape().as_list()
], output_tensor,
output_tensor.get_shape().as_list(),
[data, warp])
if not on_gpu:
# On CPU we perform numerical differentiation at the best available
# precision, and compare against that. This is necessary for test to
# pass for float16.
data_tensor_64 = constant_op.constant(data,
dtype=dtypes.float64)
warp_tensor_64 = constant_op.constant(warp,
dtype=dtypes.float64)
output_tensor_64 = resampler.resampler(data=data_tensor_64,
warp=warp_tensor_64)
grads_64 = test.compute_gradient(
[data_tensor_64, warp_tensor_64], [
data_tensor.get_shape().as_list(),
warp_tensor.get_shape().as_list()
], output_tensor_64,
output_tensor.get_shape().as_list(), [data, warp])
for g, g_64 in zip(grads, grads_64):
self.assertLess(np.fabs(g[0] - g_64[1]).max(), tol)
else:
for g in grads:
self.assertLess(np.fabs(g[0] - g[1]).max(), tol)
def _test_op_backward_pass(self, on_gpu, dtype, tol):
np.random.seed(13)
data_width = 5
data_height = 4
data_channels = 3
warp_width = 2
warp_height = 6
batch_size = 10
warp = _make_warp(batch_size, warp_height, warp_width, dtype.as_numpy_dtype)
data_shape = (batch_size, data_height, data_width, data_channels)
data = np.random.rand(*data_shape).astype(dtype.as_numpy_dtype)
with self.test_session(use_gpu=on_gpu, force_gpu=False):
data_tensor = constant_op.constant(data)
warp_tensor = constant_op.constant(warp)
output_tensor = resampler.resampler(data=data_tensor, warp=warp_tensor)
grads = test.compute_gradient([data_tensor, warp_tensor], [
data_tensor.get_shape().as_list(),
warp_tensor.get_shape().as_list()
], output_tensor, output_tensor.get_shape().as_list(), [data, warp])
if not on_gpu:
# On CPU we perform numerical differentiation at the best available
# precision, and compare against that. This is necessary for test to
# pass for float16.
data_tensor_64 = constant_op.constant(data, dtype=dtypes.float64)
warp_tensor_64 = constant_op.constant(warp, dtype=dtypes.float64)
output_tensor_64 = resampler.resampler(data=data_tensor_64,
warp=warp_tensor_64)
grads_64 = test.compute_gradient([data_tensor_64, warp_tensor_64], [
data_tensor.get_shape().as_list(),
warp_tensor.get_shape().as_list()
], output_tensor_64, output_tensor.get_shape().as_list(), [data, warp])
for g, g_64 in zip(grads, grads_64):
self.assertLess(np.fabs(g[0] - g_64[1]).max(), tol)
else:
for g in grads:
self.assertLess(np.fabs(g[0] - g[1]).max(), tol)
def _checkGradReal(self, func, x, use_gpu=False):
with self.test_session(use_gpu=use_gpu):
inx = ops.convert_to_tensor(x)
# func is a forward RFFT function (batched or unbatched).
z = func(inx)
# loss = sum(|z|^2)
loss = math_ops.reduce_sum(math_ops.real(z * math_ops.conj(z)))
x_jacob_t, x_jacob_n = test.compute_gradient(
inx, list(x.shape), loss, [1], x_init_value=x, delta=1e-2)
self.assertAllClose(x_jacob_t, x_jacob_n, rtol=1e-2, atol=1e-2)
