def _test_fn():
with backprop.GradientTape() as tape:
x = array_ops.ones([5, 5])
tape.watch(x)
y = math_ops.reduce_euclidean_norm(
x, axis=constant_op.constant(1))
return y, tape.gradient(y, x)
def test2D_4(self):
for dtype in [dtypes.float32, dtypes.float64]:
x = constant_op.constant([[3], [4]], dtype=dtype)
grads = gradient_checker_v2.compute_gradient(
lambda x: math_ops.reduce_euclidean_norm(x, 1), [x])
err = gradient_checker_v2.max_error(*grads)
self.assertLess(err, 1e-3)
def test3D_4(self):
for dtype in [dtypes.float32, dtypes.float64]:
x = constant_op.constant(
[[[-3, 5], [7, 11]], [[13, 17], [19, 23]]], dtype=dtype)
grads = gradient_checker_v2.compute_gradient(
lambda x: math_ops.reduce_euclidean_norm(x, 2), [x])
err = gradient_checker_v2.max_error(*grads)
self.assertLess(err, 2e-3)
def testComplex128(self):
for rank in range(1, _MAX_RANK + 1):
np_arr = self._makeIncremental((2, ) * rank, dtypes.complex128)
self._compareAllAxes(np_arr)
with self.session(use_gpu=True):
for dtype in (dtypes.float16, dtypes.float32, dtypes.float64):
# A large number is needed to get Eigen to die
x = array_ops.zeros((0, 9938), dtype=dtype)
y = math_ops.reduce_euclidean_norm(x, [0]).eval()
self.assertEqual(y.shape, (9938, ))
self.assertAllEqual(y, np.zeros(9938))
def testZeros(self):
for dtype in [dtypes.float32, dtypes.float64]:
x = constant_op.constant([0.0, -0.0], dtype=dtype)
with backprop.GradientTape() as tape:
tape.watch(x)
y = math_ops.reduce_euclidean_norm(x)
dx = tape.gradient(y, x)
dx_answer = constant_op.constant(
[float("NaN"), float("NaN")], dtype=dtype)
self.assertAllClose(dx, dx_answer)
def _tf_reduce(self, x, reduction_axes, keepdims):
return math_ops.reduce_euclidean_norm(x, reduction_axes, keepdims)
