def testApproximateEqual(self):
for dtype in [np.float32, np.double]:
x = dtype(1)
y = dtype(1.00009)
z = False
with test_util.device(use_gpu=True):
# Default tolerance is 0.00001
z_tf = self.evaluate(math_ops.approximate_equal(x, y))
self.assertAllEqual(z, z_tf)
for dtype in [np.float32, np.double]:
x = dtype(1)
y = dtype(1.000009)
z = True
with test_util.device(use_gpu=True):
# Default tolerance is 0.00001
z_tf = self.evaluate(math_ops.approximate_equal(x, y))
self.assertAllEqual(z, z_tf)
for dtype in [np.float32, np.double]:
x = np.array([[[[-1, 2.00009999], [-3, 4.01]]]], dtype=dtype)
y = np.array([[[[-1.001, 2], [-3.00009, 4]]]], dtype=dtype)
z = np.array([[[[False, True], [True, False]]]], dtype=np.bool)
with test_util.device(use_gpu=True):
z_tf = self.evaluate(math_ops.approximate_equal(x, y, tolerance=0.0001))
self.assertAllEqual(z, z_tf)
def testApproximateEqual(self):
for dtype in [np.float32, np.double]:
x = dtype(1)
y = dtype(1.00009)
z = False
with test_util.device(use_gpu=True):
# Default tolerance is 0.00001
z_tf = self.evaluate(math_ops.approximate_equal(x, y))
self.assertAllEqual(z, z_tf)
for dtype in [np.float32, np.double]:
x = dtype(1)
y = dtype(1.000009)
z = True
with test_util.device(use_gpu=True):
# Default tolerance is 0.00001
z_tf = self.evaluate(math_ops.approximate_equal(x, y))
self.assertAllEqual(z, z_tf)
for dtype in [np.float32, np.double]:
x = np.array([[[[-1, 2.00009999], [-3, 4.01]]]], dtype=dtype)
y = np.array([[[[-1.001, 2], [-3.00009, 4]]]], dtype=dtype)
z = np.array([[[[False, True], [True, False]]]], dtype=np.bool)
with test_util.device(use_gpu=True):
z_tf = self.evaluate(
math_ops.approximate_equal(x, y, tolerance=0.0001))
self.assertAllEqual(z, z_tf)
def testApproximateEqualShape(self):
for dtype in [np.float32, np.double]:
x = np.array([1, 2], dtype=dtype)
y = np.array([[1, 2]], dtype=dtype)
# The inputs 'x' and 'y' must have the same shape.
with self.assertRaisesRegexp(
ValueError, "Shapes must be equal rank, but are 1 and 2"):
math_ops.approximate_equal(x, y)
def testApproximateEqualShape(self):
for dtype in [np.float32, np.double]:
x = np.array([1, 2], dtype=dtype)
y = np.array([[1, 2]], dtype=dtype)
# The inputs 'x' and 'y' must have the same shape.
with self.assertRaisesRegexp(
ValueError, "Shapes must be equal rank, but are 1 and 2"):
math_ops.approximate_equal(x, y)
def testApproximateEqualShape(self):
for dtype in [np.float32, np.double]:
x = np.array([1, 2], dtype=dtype)
y = np.array([[1, 2]], dtype=dtype)
# The inputs 'x' and 'y' must have the same shape.
with self.assertRaisesRegexp(
(ValueError, errors.InvalidArgumentError),
"Shapes must be equal rank|must be of the same shape"):
math_ops.approximate_equal(x, y)
def testExecuteFloatAttr(self):
three = tensor.Tensor(3.0)
almost_three = tensor.Tensor(2.8)
almost_equal = math_ops.approximate_equal(three,
almost_three,
tolerance=0.3)
self.assertTrue(almost_equal.numpy())
def testExecuteFloatAttr(self):
three = constant_op.constant(3.0)
almost_three = constant_op.constant(2.8)
almost_equal = math_ops.approximate_equal(three,
almost_three,
tolerance=0.3)
self.assertTrue(almost_equal)
def testComplexOps(self):
for dtype in self.complex_types:
ctypes = {np.complex64: np.float32}
self._testBinary(
math_ops.complex,
np.array([[[[-1, 2], [2, 0]]]], dtype=ctypes[dtype]),
np.array([[[[2, -3], [0, 4]]]], dtype=ctypes[dtype]),
expected=np.array([[[[-1 + 2j, 2 - 3j], [2, 4j]]]], dtype=dtype))
self._testBinary(
lambda x, y: math_ops.approximate_equal(x, y, tolerance=0.0001),
np.array(
[[[[-1 + 2j, 2.00009999 - 3j], [2 - 3j, 3 + 4.01j]]]],
dtype=dtype),
np.array(
[[[[-1.001 + 2j, 2 - 3j], [2 - 3.00009j, 3 + 4j]]]], dtype=dtype),
expected=np.array([[[[False, True], [True, False]]]], dtype=dtype))
self._testBinary(
gen_math_ops._real_div,
np.array([3, 3j, -1.5j, -8, 2 + 3j, 2 + 4j, 44 + 3j], dtype=dtype),
np.array([2, -2, 7j, -4j, 4 - 6j, 1 + 2j, 0], dtype=dtype),
expected=np.array(
[
1.5, -1.5j, -0.2142857, -2j, (2 + 3j) / (4 - 6j), 2,
float("inf")
],
dtype=dtype))
# TODO(b/65408531): support+test pow for cplx
lhs = np.array([4 + 2j, -3 - 1j, 2j, 1], dtype=dtype)
rhs = np.array([5, -6j, 7 - 3j, -8j], dtype=dtype)
self._testBinary(
gen_math_ops._reciprocal_grad, lhs, rhs, expected=-rhs * lhs * lhs)
self._testBinary(
gen_math_ops._sigmoid_grad, lhs, rhs, expected=rhs * lhs * (1 - lhs))
# TODO(b/65408531): support+test _rsqrt_grad for cplx (needs pow)
self._testBinary(
gen_math_ops._sqrt_grad, lhs, rhs, expected=rhs / (2 * lhs))
self._testBinary(
gen_math_ops._tanh_grad, lhs, rhs, expected=rhs * (1 - lhs * lhs))
def testExecuteFloatAttr(self):
three = tensor.Tensor(3.0)
almost_three = tensor.Tensor(2.8)
almost_equal = math_ops.approximate_equal(
three, almost_three, tolerance=0.3)
self.assertTrue(almost_equal.numpy())
def testComplexOps(self):
for dtype in self.complex_types:
ctypes = {np.complex64: np.float32}
self._testBinary(
math_ops.complex,
np.array([[[[-1, 2], [2, 0]]]], dtype=ctypes[dtype]),
np.array([[[[2, -3], [0, 4]]]], dtype=ctypes[dtype]),
expected=np.array([[[[-1 + 2j, 2 - 3j], [2, 4j]]]], dtype=dtype))
self._testBinary(
lambda x, y: math_ops.approximate_equal(x, y, tolerance=0.0001),
np.array(
[[[[-1 + 2j, 2.00009999 - 3j], [2 - 3j, 3 + 4.01j]]]],
dtype=dtype),
np.array(
[[[[-1.001 + 2j, 2 - 3j], [2 - 3.00009j, 3 + 4j]]]], dtype=dtype),
expected=np.array([[[[False, True], [True, False]]]], dtype=dtype))
self._testBinary(
gen_math_ops._real_div,
np.array([3, 3j, -1.5j, -8, 2 + 3j, 2 + 4j], dtype=dtype),
np.array([2, -2, 7j, -4j, 4 - 6j, 1 + 2j], dtype=dtype),
expected=np.array(
[1.5, -1.5j, -0.2142857, -2j, (2 + 3j) / (4 - 6j), 2],
dtype=dtype))
# Test inf/nan scenarios.
self._testBinary(
gen_math_ops._real_div,
np.array([4 + 3j, 4, 3j, -4, -4j, 2 - 3j], dtype=dtype),
np.array([0, 0, 0, 0, 0, 0], dtype=dtype),
expected=np.array(
[
dtype(1 + 1j) / 0,
dtype(1) / 0,
dtype(1j) / 0,
dtype(-1) / 0,
dtype(-1j) / 0,
dtype(1 - 1j) / 0
],
dtype=dtype))
atan2_supported = self.device == "XLA_GPU"
if atan2_supported:
self._testBinary(
math_ops.pow,
dtype(3 + 2j),
dtype(4 - 5j),
expected=np.power(dtype(3 + 2j), dtype(4 - 5j)))
self._testBinary( # empty rhs
math_ops.pow,
np.array([1 + 2j, 2 - 3j], dtype=dtype),
np.zeros(shape=[0, 2], dtype=dtype),
expected=np.zeros(shape=[0, 2], dtype=dtype))
self._testBinary( # to zero power
math_ops.pow,
np.array([1 + 2j, 2 - 3j], dtype=dtype),
np.zeros(shape=[1, 2], dtype=dtype),
expected=np.ones(shape=[1, 2], dtype=dtype))
lhs = np.array([1 - 2j, 4 + 3j, 2 - 3j, 3, 2j, 1, 4], dtype=dtype)
rhs = np.array([2, 3j, 3 + 4j, 2 + 3j, 3 - 2j, 2, 3 + 3j], dtype=dtype)
scalar = dtype(2 + 2j)
self._testBinary(math_ops.pow, lhs, rhs, expected=np.power(lhs, rhs))
self._testBinary(
math_ops.pow, scalar, rhs, expected=np.power(scalar, rhs))
self._testBinary(math_ops.pow, lhs, scalar, np.power(lhs, scalar))
lhs = np.array([4 + 2j, -3 - 1j, 2j, 1], dtype=dtype)
rhs = np.array([5, -6j, 7 - 3j, -8j], dtype=dtype)
self._testBinary(
gen_math_ops._reciprocal_grad, lhs, rhs, expected=-rhs * lhs * lhs)
self._testBinary(
gen_math_ops._sigmoid_grad, lhs, rhs, expected=rhs * lhs * (1 - lhs))
if atan2_supported:
self._testBinary(
gen_math_ops._rsqrt_grad, lhs, rhs, expected=lhs**3 * rhs / -2)
self._testBinary(
gen_math_ops._sqrt_grad, lhs, rhs, expected=rhs / (2 * lhs))
self._testBinary(
gen_math_ops._tanh_grad, lhs, rhs, expected=rhs * (1 - lhs * lhs))
def testComplexOps(self):
for dtype in self.complex_types:
ctypes = {np.complex64: np.float32}
self._testBinary(
math_ops.complex,
np.array([[[[-1, 2], [2, 0]]]], dtype=ctypes[dtype]),
np.array([[[[2, -3], [0, 4]]]], dtype=ctypes[dtype]),
expected=np.array([[[[-1 + 2j, 2 - 3j], [2, 4j]]]], dtype=dtype))
self._testBinary(
lambda x, y: math_ops.approximate_equal(x, y, tolerance=0.0001),
np.array(
[[[[-1 + 2j, 2.00009999 - 3j], [2 - 3j, 3 + 4.01j]]]],
dtype=dtype),
np.array(
[[[[-1.001 + 2j, 2 - 3j], [2 - 3.00009j, 3 + 4j]]]], dtype=dtype),
expected=np.array([[[[False, True], [True, False]]]], dtype=dtype))
self._testBinary(
gen_math_ops._real_div,
np.array([3, 3j, -1.5j, -8, 2 + 3j, 2 + 4j], dtype=dtype),
np.array([2, -2, 7j, -4j, 4 - 6j, 1 + 2j], dtype=dtype),
expected=np.array(
[1.5, -1.5j, -0.2142857, -2j, (2 + 3j) / (4 - 6j), 2],
dtype=dtype))
# Test inf/nan scenarios.
self._testBinary(
gen_math_ops._real_div,
np.array([4 + 3j, 4, 3j, -4, -4j, 2 - 3j], dtype=dtype),
np.array([0, 0, 0, 0, 0, 0], dtype=dtype),
expected=np.array(
[
dtype(1 + 1j) / 0,
dtype(1) / 0,
dtype(1j) / 0,
dtype(-1) / 0,
dtype(-1j) / 0,
dtype(1 - 1j) / 0
],
dtype=dtype))
self._testBinary(
math_ops.pow,
dtype(3 + 2j),
dtype(4 - 5j),
expected=np.power(dtype(3 + 2j), dtype(4 - 5j)))
self._testBinary( # empty rhs
math_ops.pow,
np.array([1 + 2j, 2 - 3j], dtype=dtype),
np.zeros(shape=[0, 2], dtype=dtype),
expected=np.zeros(shape=[0, 2], dtype=dtype))
self._testBinary( # to zero power
math_ops.pow,
np.array([1 + 2j, 2 - 3j], dtype=dtype),
np.zeros(shape=[1, 2], dtype=dtype),
expected=np.ones(shape=[1, 2], dtype=dtype))
lhs = np.array([1 - 2j, 4 + 3j, 2 - 3j, 3, 2j, 1, 4], dtype=dtype)
rhs = np.array([2, 3j, 3 + 4j, 2 + 3j, 3 - 2j, 2, 3 + 3j], dtype=dtype)
scalar = dtype(2 + 2j)
self._testBinary(math_ops.pow, lhs, rhs, expected=np.power(lhs, rhs))
self._testBinary(
math_ops.pow, scalar, rhs, expected=np.power(scalar, rhs))
self._testBinary(math_ops.pow, lhs, scalar, np.power(lhs, scalar))
lhs = np.array([4 + 2j, -3 - 1j, 2j, 1], dtype=dtype)
rhs = np.array([5, -6j, 7 - 3j, -8j], dtype=dtype)
self._testBinary(
gen_math_ops._reciprocal_grad, lhs, rhs, expected=-rhs * lhs * lhs)
self._testBinary(
gen_math_ops._sigmoid_grad, lhs, rhs, expected=rhs * lhs * (1 - lhs))
self._testBinary(
gen_math_ops._rsqrt_grad, lhs, rhs, expected=lhs**3 * rhs / -2)
self._testBinary(
gen_math_ops._sqrt_grad, lhs, rhs, expected=rhs / (2 * lhs))
self._testBinary(
gen_math_ops._tanh_grad, lhs, rhs, expected=rhs * (1 - lhs * lhs))
def approximate_equal(x, y): return math_ops.approximate_equal(x, y)
def testFloatOps(self):
for dtype in self.float_types:
self._testBinary(lambda x, y: math_ops.approximate_equal(
x, y, tolerance=0.0001),
np.array([[[[-1, 2.00009999], [-3, 4.01]]]],
dtype=dtype),
np.array([[[[-1.001, 2], [-3.00009, 4]]]],
dtype=dtype),
expected=np.array(
[[[[False, True], [True, False]]]],
dtype=dtype))
self._testBinary(gen_math_ops._real_div,
np.array([3, 3, -1.5, -8, 44], dtype=dtype),
np.array([2, -2, 7, -4, 0], dtype=dtype),
expected=np.array(
[1.5, -1.5, -0.2142857, 2,
float("inf")],
dtype=dtype))
self._testBinary(math_ops.pow,
dtype(3),
dtype(4),
expected=dtype(81))
self._testBinary(math_ops.pow,
np.array([1, 2], dtype=dtype),
np.zeros(shape=[0, 2], dtype=dtype),
expected=np.zeros(shape=[0, 2], dtype=dtype))
self._testBinary(math_ops.pow,
np.array([10, 4], dtype=dtype),
np.array([2, 3], dtype=dtype),
expected=np.array([100, 64], dtype=dtype))
self._testBinary(math_ops.pow,
dtype(2),
np.array([3, 4], dtype=dtype),
expected=np.array([8, 16], dtype=dtype))
self._testBinary(math_ops.pow,
np.array([[2], [3]], dtype=dtype),
dtype(4),
expected=np.array([[16], [81]], dtype=dtype))
self._testBinary(gen_math_ops._reciprocal_grad,
np.array([4, -3, -2, 1], dtype=dtype),
np.array([5, -6, 7, -8], dtype=dtype),
expected=np.array([-80, 54, -28, 8], dtype=dtype))
self._testBinary(gen_math_ops._sigmoid_grad,
np.array([4, 3, 2, 1], dtype=dtype),
np.array([5, 6, 7, 8], dtype=dtype),
expected=np.array([-60, -36, -14, 0],
dtype=dtype))
self._testBinary(gen_math_ops._rsqrt_grad,
np.array([4, 3, 2, 1], dtype=dtype),
np.array([5, 6, 7, 8], dtype=dtype),
expected=np.array([-160, -81, -28, -4],
dtype=dtype))
self._testBinary(gen_math_ops._sqrt_grad,
np.array([4, 3, 2, 1], dtype=dtype),
np.array([5, 6, 7, 8], dtype=dtype),
expected=np.array([0.625, 1, 1.75, 4],
dtype=dtype))
self._testBinary(
gen_nn_ops._softplus_grad,
np.array([4, 3, 2, 1], dtype=dtype),
np.array([5, 6, 7, 8], dtype=dtype),
expected=np.array(
[3.97322869, 2.99258232, 1.99817801, 0.99966466],
dtype=dtype))
self._testBinary(gen_nn_ops._softsign_grad,
np.array([4, 3, 2, 1], dtype=dtype),
np.array([5, 6, 7, 8], dtype=dtype),
expected=np.array(
[0.11111111, 0.06122449, 0.03125, 0.01234568],
dtype=dtype))
self._testBinary(gen_math_ops._tanh_grad,
np.array([4, 3, 2, 1], dtype=dtype),
np.array([5, 6, 7, 8], dtype=dtype),
expected=np.array([-75, -48, -21, 0],
dtype=dtype))
self._testBinary(gen_nn_ops._elu_grad,
np.array([1, 2, 3, 4, 5, 6], dtype=dtype),
np.array([-.6, -.4, -.2, 0, .2, .4], dtype=dtype),
expected=np.array([0.4, 1.2, 2.4, 4, 5, 6],
dtype=dtype))
self._testBinary(
gen_nn_ops._selu_grad,
np.array([1, 2, 3, 4, 5, 6], dtype=dtype),
np.array([-.6, -.4, -.2, .2, .4, .6], dtype=dtype),
expected=np.array([
1.158099340847, 2.7161986816948, 4.67429802254,
4.202803949422, 5.2535049367774, 6.30420592413
],
dtype=dtype))
self._testBinary(gen_nn_ops._relu_grad,
np.array([1, 2, 3, 4, 5, 6, 7, 8, 9, 10],
dtype=dtype),
np.array([0, 0, 0, 0, 0, 0.1, 0.3, 0.5, 0.7, 0.9],
dtype=dtype),
expected=np.array([0, 0, 0, 0, 0, 6, 7, 8, 9, 10],
dtype=dtype))
self._testBinary(
gen_nn_ops._relu6_grad,
np.array([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12], dtype=dtype),
np.array([0, 0, 0, 0, 0, 0.1, 0.3, 0.5, 0.7, 0.9, 6.1, 10.0],
dtype=dtype),
expected=np.array([0, 0, 0, 0, 0, 6, 7, 8, 9, 10, 0, 0],
dtype=dtype))
self._testBinary(
gen_nn_ops._softmax_cross_entropy_with_logits,
np.array([[1, 2, 3, 4], [5, 6, 7, 8]], dtype=dtype),
np.array([[0.1, 0.2, 0.3, 0.4], [0.4, 0.3, 0.2, 0.1]],
dtype=dtype),
expected=[
np.array([1.44019, 2.44019], dtype=dtype),
np.array([[-0.067941, -0.112856, -0.063117, 0.243914],
[-0.367941, -0.212856, 0.036883, 0.543914]],
dtype=dtype),
],
equality_test=self.ListsAreClose)
self._testBinary(
gen_nn_ops._sparse_softmax_cross_entropy_with_logits,
np.array([[0.1, 0.2, 0.3, 0.4], [0.5, 0.6, 0.7, 0.8],
[0.9, 1.0, 1.1, 1.2]],
dtype=dtype),
np.array([2, 1, 7], dtype=np.int32),
expected=[
np.array([1.342536, 1.442536, np.nan], dtype=dtype),
np.array([[0.213838, 0.236328, -0.738817, 0.288651],
[0.213838, -0.763672, 0.261183, 0.288651],
[np.nan, np.nan, np.nan, np.nan]],
dtype=dtype),
],
equality_test=self.ListsAreClose)
def loop_fn(i): x1 = array_ops.gather(x, i) y1 = array_ops.gather(y, i) return math_ops.approximate_equal(x1, y1)
def testExecuteFloatAttr(self):
three = constant_op.constant(3.0)
almost_three = constant_op.constant(2.8)
almost_equal = math_ops.approximate_equal(
three, almost_three, tolerance=0.3)
self.assertTrue(almost_equal)
def loop_fn(i):
x1 = array_ops.gather(x, i)
y1 = array_ops.gather(y, i)
return math_ops.approximate_equal(x1, y1)
def testFloatOps(self):
for dtype in self.float_types:
if dtype == dtypes.bfloat16.as_numpy_dtype:
a = -1.01
b = 4.1
else:
a = -1.001
b = 4.01
self._testBinary(
lambda x, y: math_ops.approximate_equal(x, y, tolerance=0.0001),
np.array([[[[-1, 2.00009999], [-3, b]]]], dtype=dtype),
np.array([[[[a, 2], [-3.00009, 4]]]], dtype=dtype),
expected=np.array([[[[False, True], [True, False]]]], dtype=dtype))
self._testBinary(
gen_math_ops._real_div,
np.array([3, 3, -1.5, -8, 44], dtype=dtype),
np.array([2, -2, 7, -4, 0], dtype=dtype),
expected=np.array(
[1.5, -1.5, -0.2142857, 2, float("inf")], dtype=dtype))
self._testBinary(math_ops.pow, dtype(3), dtype(4), expected=dtype(81))
self._testBinary(
math_ops.pow,
np.array([1, 2], dtype=dtype),
np.zeros(shape=[0, 2], dtype=dtype),
expected=np.zeros(shape=[0, 2], dtype=dtype))
self._testBinary(
math_ops.pow,
np.array([10, 4], dtype=dtype),
np.array([2, 3], dtype=dtype),
expected=np.array([100, 64], dtype=dtype))
self._testBinary(
math_ops.pow,
dtype(2),
np.array([3, 4], dtype=dtype),
expected=np.array([8, 16], dtype=dtype))
self._testBinary(
math_ops.pow,
np.array([[2], [3]], dtype=dtype),
dtype(4),
expected=np.array([[16], [81]], dtype=dtype))
self._testBinary(
math_ops.atan2,
np.array([0, np.sqrt(2), 1, np.sqrt(2), 0], dtype),
np.array([1, np.sqrt(2), 0, -np.sqrt(2), -1], dtype),
expected=np.array(
[0, np.pi / 4, np.pi / 2, np.pi * 3 / 4, np.pi], dtype=dtype))
self._testBinary(
gen_math_ops._reciprocal_grad,
np.array([4, -3, -2, 1], dtype=dtype),
np.array([5, -6, 7, -8], dtype=dtype),
expected=np.array([-80, 54, -28, 8], dtype=dtype))
self._testBinary(
gen_math_ops._sigmoid_grad,
np.array([4, 3, 2, 1], dtype=dtype),
np.array([5, 6, 7, 8], dtype=dtype),
expected=np.array([-60, -36, -14, 0], dtype=dtype))
self._testBinary(
gen_math_ops._rsqrt_grad,
np.array([4, 3, 2, 1], dtype=dtype),
np.array([5, 6, 7, 8], dtype=dtype),
expected=np.array([-160, -81, -28, -4], dtype=dtype))
self._testBinary(
gen_math_ops._sqrt_grad,
np.array([4, 3, 2, 1], dtype=dtype),
np.array([5, 6, 7, 8], dtype=dtype),
expected=np.array([0.625, 1, 1.75, 4], dtype=dtype))
self._testBinary(
gen_nn_ops._softplus_grad,
np.array([4, 3, 2, 1], dtype=dtype),
np.array([5, 6, 7, 8], dtype=dtype),
expected=np.array(
[3.97322869, 2.99258232, 1.99817801, 0.99966466], dtype=dtype))
self._testBinary(
gen_nn_ops._softsign_grad,
np.array([4, 3, 2, 1], dtype=dtype),
np.array([5, 6, 7, 8], dtype=dtype),
expected=np.array(
[0.11111111, 0.06122449, 0.03125, 0.01234568], dtype=dtype))
self._testBinary(
gen_math_ops._tanh_grad,
np.array([4, 3, 2, 1], dtype=dtype),
np.array([5, 6, 7, 8], dtype=dtype),
expected=np.array([-75, -48, -21, 0], dtype=dtype))
self._testBinary(
gen_nn_ops._elu_grad,
np.array([1, 2, 3, 4, 5, 6], dtype=dtype),
np.array([-.6, -.4, -.2, 0, .2, .4], dtype=dtype),
expected=np.array([0.4, 1.2, 2.4, 4, 5, 6], dtype=dtype))
self._testBinary(
gen_nn_ops._selu_grad,
np.array([1, 2, 3, 4, 5, 6], dtype=dtype),
np.array([-.6, -.4, -.2, .2, .4, .6], dtype=dtype),
expected=np.array(
[1.158099340847, 2.7161986816948, 4.67429802254,
4.202803949422, 5.2535049367774, 6.30420592413], dtype=dtype))
self._testBinary(
gen_nn_ops._relu_grad,
np.array([1, 2, 3, 4, 5, 6, 7, 8, 9, 10], dtype=dtype),
np.array([0, 0, 0, 0, 0, 0.1, 0.3, 0.5, 0.7, 0.9], dtype=dtype),
expected=np.array([0, 0, 0, 0, 0, 6, 7, 8, 9, 10], dtype=dtype))
self._testBinary(
gen_nn_ops._relu6_grad,
np.array([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12], dtype=dtype),
np.array(
[0, 0, 0, 0, 0, 0.1, 0.3, 0.5, 0.7, 0.9, 6.1, 10.0], dtype=dtype),
expected=np.array([0, 0, 0, 0, 0, 6, 7, 8, 9, 10, 0, 0], dtype=dtype))
self._testBinary(
gen_nn_ops._softmax_cross_entropy_with_logits,
np.array([[1, 2, 3, 4], [5, 6, 7, 8]], dtype=dtype),
np.array([[0.1, 0.2, 0.3, 0.4], [0.4, 0.3, 0.2, 0.1]], dtype=dtype),
expected=[
np.array([1.44019, 2.44019], dtype=dtype),
np.array([[-0.067941, -0.112856, -0.063117, 0.243914],
[-0.367941, -0.212856, 0.036883, 0.543914]],
dtype=dtype),
],
equality_test=self.ListsAreClose)
self._testBinary(
gen_nn_ops._sparse_softmax_cross_entropy_with_logits,
np.array([[0.1, 0.2, 0.3, 0.4], [0.5, 0.6, 0.7, 0.8],
[0.9, 1.0, 1.1, 1.2]], dtype=dtype),
np.array([2, 1, 7], dtype=np.int32),
expected=[
np.array([1.342536, 1.442536, np.nan], dtype=dtype),
np.array([[0.213838, 0.236328, -0.738817, 0.288651],
[0.213838, -0.763672, 0.261183, 0.288651],
[np.nan, np.nan, np.nan, np.nan]],
dtype=dtype),
],
equality_test=self.ListsAreClose)
