def quantize_and_dequantize_v2_round_half_to_even(x):
return array_ops.quantize_and_dequantize_v2(
x,
-1.0,
1.0,
signed_input=True,
num_bits=8,
range_given=True,
round_mode="HALF_TO_EVEN")
def quantize_and_dequantize_v2_round_half_to_even(x):
return array_ops.quantize_and_dequantize_v2(
x,
-1.0,
1.0,
signed_input=True,
num_bits=8,
range_given=True,
round_mode="HALF_TO_EVEN")
def quantize_and_dequantize_v2_round_half_up(x):
return array_ops.quantize_and_dequantize_v2(
x,
-1,
1.0,
signed_input=True,
num_bits=8,
range_given=True,
round_mode="HALF_UP")
def quantize_and_dequantize_v2_round_half_up(x):
return array_ops.quantize_and_dequantize_v2(
x,
-1,
1.0,
signed_input=True,
num_bits=8,
range_given=True,
round_mode="HALF_UP")
def testFloatOps(self):
for dtype in self.float_types:
self._assertOpOutputMatchesExpected(
math_ops.acosh,
np.array([1, 2, 3, 4], dtype=dtype),
expected=np.array([0, 1.3169579, 1.76274717, 2.06343707],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.asinh,
np.array([1, 2, 3, 4], dtype=dtype),
expected=np.array([0.88137359, 1.44363548, 1.81844646, 2.09471255],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.atanh,
np.array([0.1, 0.2, 0.3, 0.4], dtype=dtype),
expected=np.array([0.10033535, 0.20273255, 0.3095196, 0.42364893],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.ceil,
np.array([[-1.7, 1.2]], dtype=dtype),
expected=np.array([[-1, 2]], dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.cosh,
np.array([1, 2, 3, 4], dtype=dtype),
expected=np.array([1.54308063, 3.76219569, 10.067662, 27.30823284],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.exp,
np.array([[-1, 1]], dtype=dtype),
expected=np.array([[0.36787945, 2.7182817]], dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.expm1,
np.array([[-1, 1]], dtype=dtype),
expected=np.array([[-0.63212056, 1.71828183]], dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.floor,
np.array([[-1.7, 1.2]], dtype=dtype),
expected=np.array([[-2, 1]], dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.is_finite,
np.array([[np.NINF, -2, -1, 0, 0.5, 1, 2, np.inf, np.nan]],
dtype=dtype),
expected=np.array([[0, 1, 1, 1, 1, 1, 1, 0, 0]], dtype=np.bool))
# Tests for tf.nn ops.
self._assertOpOutputMatchesExpected(
nn_ops.l2_loss, np.array([[[]]], dtype=dtype), expected=dtype(0))
self._assertOpOutputMatchesExpected(nn_ops.l2_loss, dtype(4), dtype(8))
self._assertOpOutputMatchesExpected(
nn_ops.l2_loss, np.array([[-2, 4]], dtype=dtype), expected=dtype(10))
self._assertOpOutputMatchesExpected(
math_ops.reciprocal,
np.array([[1, 2]], dtype=dtype),
expected=np.array([[1, 0.5]], dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.log,
np.array([[1, 2]], dtype=dtype),
expected=np.array([[0, 0.69314718]], dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.sin,
np.array([[1, 2]], dtype=dtype),
expected=np.array([[0.841478, 0.909302]], dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.cos,
np.array([[1, 2]], dtype=dtype),
expected=np.array([[0.540297, -0.41614]], dtype=dtype))
# TODO(b/34703906): improve log1p implementation and make tolerance
# tighter.
self._assertOpOutputMatchesExpected(
math_ops.log1p,
np.array([[1e-14, 1e-15, 0.6]], dtype=dtype),
expected=np.log1p(np.array([[1e-14, 1e-15, 0.6]], dtype=dtype)))
self._assertOpOutputMatchesExpected(
math_ops.rint,
np.array([[-1.7, 1.2, 4.0, 0.0], [-3.5, -2.5, -1.5, -0.5],
[0.5, 1.5, 2.5, 3.5]], dtype=dtype),
expected=np.array([[-2, 1, 4, 0], [-4, -2, -2, 0], [0, 2, 2, 4]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.round,
np.array([[-1.7, 1.2, 4.0, 0.0], [-3.5, -2.5, -1.5, -0.5],
[0.5, 1.5, 2.5, 3.5]], dtype=dtype),
expected=np.array([[-2, 1, 4, 0], [-4, -2, -2, 0], [0, 2, 2, 4]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.rsqrt,
np.array([[4, 16]], dtype=dtype),
expected=np.array([[0.5, 0.25]], dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.sigmoid,
np.array(
[[1, 1, 1, 1],
[1, 2, 3, 4]],
dtype=dtype),
expected=np.array(
[[0.7310586, 0.7310586, 0.7310586, 0.7310586],
[0.7310586, 0.880797, 0.95257413, 0.98201376]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.sigmoid,
np.array([-300, -150, 0, 150, 300], dtype=dtype),
expected=np.array([0, 0, 0.5, 1, 1], dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.sinh,
np.array([1, 2, 3, 4], dtype=dtype),
expected=np.array([1.17520119, 3.62686041, 10.01787493, 27.2899172],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.sqrt,
np.array([[4, 9]], dtype=dtype),
expected=np.array([[2, 3]], dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.tan,
np.array([1, 2, 3, 4], dtype=dtype),
expected=np.array([1.55740772, -2.18503986, -0.14254654, 1.15782128],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.tanh,
np.array(
[[1, 1, 1, 1],
[1, 2, 3, 4]],
dtype=dtype),
expected=np.array(
[[0.76159418, 0.76159418, 0.76159418, 0.76159418],
[0.76159418, 0.96402758, 0.99505478, 0.99932933]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
nn_ops.log_softmax,
np.array(
[[1, 1, 1, 1],
[1, 2, 3, 4]],
dtype=dtype),
expected=np.array(
[[-1.3862944, -1.3862944, -1.3862944, -1.3862944],
[-3.4401896, -2.4401896, -1.4401897, -0.44018969]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
nn_ops.elu,
np.array([[-1, 0, 1]], dtype=dtype),
expected=np.array([[-0.63212056, 0, 1]], dtype=dtype))
self._assertOpOutputMatchesExpected(
nn_ops.selu,
np.array([[-1, 0, 1]], dtype=dtype),
expected=np.array([[-1.11133074, 0., 1.05070099]], dtype=dtype))
self._assertOpOutputMatchesExpected(
nn_ops.relu,
np.array([[-1, 1]], dtype=dtype),
expected=np.array([[0, 1]], dtype=dtype))
self._assertOpOutputMatchesExpected(
nn_ops.relu6,
np.array([[-0.05, 6.05, 5]], dtype=dtype),
expected=np.array([[0, 6, 5]], dtype=dtype))
self._assertOpOutputMatchesExpected(
nn_ops.softmax,
np.array(
[[1, 1, 1, 1],
[1, 2, 3, 4]],
dtype=dtype),
expected=np.array(
[[0.25, 0.25, 0.25, 0.25],
[0.032058604, 0.087144323, 0.23688284, 0.64391428]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
nn_ops.softsign,
np.array([[-2, -1, 0, 1, 2]], dtype=dtype),
expected=np.array([[-0.66666669, -0.5, 0, 0.5, 0.66666669]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.is_finite,
np.array(
[[42, float("inf"), -123], [float("nan"), 0, -0.0]], dtype=dtype),
expected=np.array(
[[True, False, True], [False, True, True]], dtype=np.bool))
self._assertOpOutputMatchesExpected(
lambda x: array_ops.quantize_and_dequantize_v2(x, -127, 127, True, 8),
np.array([-1, -0.5, 0, 0.3], dtype=dtype),
expected=np.array([-1, -64.0 / 127, 0, 38.0 / 127], dtype=dtype))
def quantize_and_dequantize_v2(x):
return array_ops.quantize_and_dequantize_v2(
x, -127, 127, signed_input=True, num_bits=8)
def quantize_and_dequantize_v2(x):
return array_ops.quantize_and_dequantize_v2(
x, -127, 127, signed_input=True, num_bits=8)
def testFloatOps(self):
for dtype in self.float_types:
self._assertOpOutputMatchesExpected(
math_ops.acosh,
np.array([1, 2, 3, 4], dtype=dtype),
expected=np.array([0, 1.3169579, 1.76274717, 2.06343707],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.asinh,
np.array([1, 2, 3, 4], dtype=dtype),
expected=np.array([0.88137359, 1.44363548, 1.81844646, 2.09471255],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.atanh,
np.array([0.1, 0.2, 0.3, 0.4], dtype=dtype),
expected=np.array([0.10033535, 0.20273255, 0.3095196, 0.42364893],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.ceil,
np.array([[-1.7, 1.2]], dtype=dtype),
expected=np.array([[-1, 2]], dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.cosh,
np.array([1, 2, 3, 4], dtype=dtype),
expected=np.array([1.54308063, 3.76219569, 10.067662, 27.30823284],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.exp,
np.array([[-1, 1]], dtype=dtype),
expected=np.array([[0.36787945, 2.7182817]], dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.expm1,
np.array([[-1, 1]], dtype=dtype),
expected=np.array([[-0.63212056, 1.71828183]], dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.floor,
np.array([[-1.7, 1.2]], dtype=dtype),
expected=np.array([[-2, 1]], dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.is_finite,
np.array([[np.NINF, -2, -1, 0, 0.5, 1, 2, np.inf, np.nan]],
dtype=dtype),
expected=np.array([[0, 1, 1, 1, 1, 1, 1, 0, 0]], dtype=np.bool))
# Tests for tf.nn ops.
self._assertOpOutputMatchesExpected(
nn_ops.l2_loss, np.array([[[]]], dtype=dtype), expected=dtype(0))
self._assertOpOutputMatchesExpected(nn_ops.l2_loss, dtype(4), dtype(8))
self._assertOpOutputMatchesExpected(
nn_ops.l2_loss, np.array([[-2, 4]], dtype=dtype), expected=dtype(10))
self._assertOpOutputMatchesExpected(
math_ops.reciprocal,
np.array([[1, 2]], dtype=dtype),
expected=np.array([[1, 0.5]], dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.log,
np.array([[1, 2]], dtype=dtype),
expected=np.array([[0, 0.69314718]], dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.sin,
np.array([[1, 2]], dtype=dtype),
expected=np.array([[0.841478, 0.909302]], dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.cos,
np.array([[1, 2]], dtype=dtype),
expected=np.array([[0.540297, -0.41614]], dtype=dtype))
# TODO(b/34703906): improve log1p implementation and make tolerance
# tighter.
self._assertOpOutputMatchesExpected(
math_ops.log1p,
np.array([[1e-14, 1e-15, 0.6]], dtype=dtype),
expected=np.log1p(np.array([[1e-14, 1e-15, 0.6]], dtype=dtype)))
self._assertOpOutputMatchesExpected(
math_ops.rint,
np.array([[-1.7, 1.2, 4.0, 0.0], [-3.5, -2.5, -1.5, -0.5],
[0.5, 1.5, 2.5, 3.5]], dtype=dtype),
expected=np.array([[-2, 1, 4, 0], [-4, -2, -2, 0], [0, 2, 2, 4]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.round,
np.array([[-1.7, 1.2, 4.0, 0.0], [-3.5, -2.5, -1.5, -0.5],
[0.5, 1.5, 2.5, 3.5]], dtype=dtype),
expected=np.array([[-2, 1, 4, 0], [-4, -2, -2, 0], [0, 2, 2, 4]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.rsqrt,
np.array([[4, 16]], dtype=dtype),
expected=np.array([[0.5, 0.25]], dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.sigmoid,
np.array(
[[1, 1, 1, 1],
[1, 2, 3, 4]],
dtype=dtype),
expected=np.array(
[[0.7310586, 0.7310586, 0.7310586, 0.7310586],
[0.7310586, 0.880797, 0.95257413, 0.98201376]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.sigmoid,
np.array([-300, -150, 0, 150, 300], dtype=dtype),
expected=np.array([0, 0, 0.5, 1, 1], dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.sinh,
np.array([1, 2, 3, 4], dtype=dtype),
expected=np.array([1.17520119, 3.62686041, 10.01787493, 27.2899172],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.sqrt,
np.array([[4, 9]], dtype=dtype),
expected=np.array([[2, 3]], dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.tan,
np.array([1, 2, 3, 4], dtype=dtype),
expected=np.array([1.55740772, -2.18503986, -0.14254654, 1.15782128],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.tanh,
np.array(
[[1, 1, 1, 1],
[1, 2, 3, 4]],
dtype=dtype),
expected=np.array(
[[0.76159418, 0.76159418, 0.76159418, 0.76159418],
[0.76159418, 0.96402758, 0.99505478, 0.99932933]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
nn_ops.log_softmax,
np.array(
[[1, 1, 1, 1],
[1, 2, 3, 4]],
dtype=dtype),
expected=np.array(
[[-1.3862944, -1.3862944, -1.3862944, -1.3862944],
[-3.4401896, -2.4401896, -1.4401897, -0.44018969]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
nn_ops.elu,
np.array([[-1, 0, 1]], dtype=dtype),
expected=np.array([[-0.63212056, 0, 1]], dtype=dtype))
self._assertOpOutputMatchesExpected(
nn_ops.selu,
np.array([[-1, 0, 1]], dtype=dtype),
expected=np.array([[-1.11133074, 0., 1.05070099]], dtype=dtype))
self._assertOpOutputMatchesExpected(
nn_ops.relu,
np.array([[-1, 1]], dtype=dtype),
expected=np.array([[0, 1]], dtype=dtype))
self._assertOpOutputMatchesExpected(
nn_ops.relu6,
np.array([[-0.05, 6.05, 5]], dtype=dtype),
expected=np.array([[0, 6, 5]], dtype=dtype))
self._assertOpOutputMatchesExpected(
nn_ops.softmax,
np.array(
[[1, 1, 1, 1],
[1, 2, 3, 4]],
dtype=dtype),
expected=np.array(
[[0.25, 0.25, 0.25, 0.25],
[0.032058604, 0.087144323, 0.23688284, 0.64391428]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
nn_ops.softsign,
np.array([[-2, -1, 0, 1, 2]], dtype=dtype),
expected=np.array([[-0.66666669, -0.5, 0, 0.5, 0.66666669]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.is_finite,
np.array(
[[42, float("inf"), -123], [float("nan"), 0, -0.0]], dtype=dtype),
expected=np.array(
[[True, False, True], [False, True, True]], dtype=np.bool))
self._assertOpOutputMatchesExpected(
lambda x: array_ops.quantize_and_dequantize_v2(x, -127, 127, True, 8),
np.array([-1, -0.5, 0, 0.3], dtype=dtype),
expected=np.array([-1, -64.0 / 127, 0, 38.0 / 127], dtype=dtype))
def testFloatOps(self):
for dtype in self.float_types:
# TODO(b/77694432): Half test failed on CPU, last ran on 04-06-2018.
if dtype == np.float16 and self.device == "XLA_CPU":
continue
x = np.arange(-0.90, 0.90, 0.25)
self._assertOpOutputMatchesExpected(
math_ops.acos,
x.astype(dtype),
expected=np.arccos(x).astype(dtype))
self._assertOpOutputMatchesExpected(
math_ops.asin,
x.astype(dtype),
expected=np.arcsin(x).astype(dtype))
x = np.arange(-3, 3).reshape(1, 3, 2)
self._assertOpOutputMatchesExpected(
math_ops.atan,
x.astype(dtype),
expected=np.arctan(x).astype(dtype))
self._assertOpOutputMatchesExpected(
math_ops.acosh,
np.array([1, 2, 3, 4], dtype=dtype),
expected=np.array([0, 1.3169579, 1.76274717, 2.06343707],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.asinh,
np.array([1, 2, 3, 4], dtype=dtype),
expected=np.array(
[0.88137359, 1.44363548, 1.81844646, 2.09471255],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.atanh,
np.array([0.1, 0.2, 0.3, 0.4], dtype=dtype),
expected=np.array(
[0.10033535, 0.20273255, 0.3095196, 0.42364893],
dtype=dtype))
self._assertOpOutputMatchesExpected(math_ops.ceil,
np.array([[-1.7, 1.2]],
dtype=dtype),
expected=np.array([[-1, 2]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.cosh,
np.array([1, 2, 3, 4], dtype=dtype),
expected=np.array(
[1.54308063, 3.76219569, 10.067662, 27.30823284],
dtype=dtype))
# Disable float16 testing for now
if dtype != np.float16:
x = np.arange(-10, 10, 1).astype(dtype)
with self.test_session() as session:
erf_x = session.run(math_ops.erf(x))
erfc_x = session.run(math_ops.erfc(x))
self._assertOpOutputMatchesExpected(math_ops.erf,
x,
expected=erf_x)
self._assertOpOutputMatchesExpected(math_ops.erfc,
x,
expected=erfc_x)
self._assertOpOutputMatchesExpected(math_ops.exp,
np.array([[-1, 1]],
dtype=dtype),
expected=np.array(
[[0.36787945, 2.7182817]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.expm1,
np.array([[-1, 1]], dtype=dtype),
expected=np.array([[-0.63212056, 1.71828183]], dtype=dtype),
rtol=1e-5)
self._assertOpOutputMatchesExpected(math_ops.floor,
np.array([[-1.7, 1.2]],
dtype=dtype),
expected=np.array([[-2, 1]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.is_finite,
np.array([[np.NINF, -2, -1, 0, 0.5, 1, 2, np.inf, np.nan]],
dtype=dtype),
expected=np.array([[0, 1, 1, 1, 1, 1, 1, 0, 0]],
dtype=np.bool))
# Tests for tf.nn ops.
self._assertOpOutputMatchesExpected(nn_ops.l2_loss,
np.array([[[]]], dtype=dtype),
expected=dtype(0))
self._assertOpOutputMatchesExpected(nn_ops.l2_loss, dtype(4),
dtype(8))
self._assertOpOutputMatchesExpected(nn_ops.l2_loss,
np.array([[-2, 4]],
dtype=dtype),
expected=dtype(10))
self._assertOpOutputMatchesExpected(math_ops.reciprocal,
np.array([[1, 2]],
dtype=dtype),
expected=np.array([[1, 0.5]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.log,
np.array([[1, 2]], dtype=dtype),
expected=np.array([[0, 0.69314718]], dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.sin,
np.array([[1, 2]], dtype=dtype),
expected=np.array([[0.841478, 0.909302]], dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.cos,
np.array([[1, 2]], dtype=dtype),
expected=np.array([[0.540297, -0.41614]], dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.log1p,
np.array([[1e-14, 1e-15, 0.6]], dtype=dtype),
expected=np.log1p(np.array([[1e-14, 1e-15, 0.6]],
dtype=dtype)),
rtol=1e-4,
atol=1e-6)
self._assertOpOutputMatchesExpected(
math_ops.rint,
np.array([[-1.7, 1.2, 4.0, 0.0], [-3.5, -2.5, -1.5, -0.5],
[0.5, 1.5, 2.5, 3.5]],
dtype=dtype),
expected=np.array(
[[-2, 1, 4, 0], [-4, -2, -2, 0], [0, 2, 2, 4]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.round,
np.array([[-1.7, 1.2, 4.0, 0.0], [-3.5, -2.5, -1.5, -0.5],
[0.5, 1.5, 2.5, 3.5]],
dtype=dtype),
expected=np.array(
[[-2, 1, 4, 0], [-4, -2, -2, 0], [0, 2, 2, 4]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.rsqrt,
np.array([[4, 16]], dtype=dtype),
expected=np.array([[0.5, 0.25]], dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.sigmoid,
np.array([[1, 1, 1, 1], [1, 2, 3, 4]], dtype=dtype),
expected=np.array(
[[0.7310586, 0.7310586, 0.7310586, 0.7310586],
[0.7310586, 0.880797, 0.95257413, 0.98201376]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.sigmoid,
np.array([-300, -150, 0, 150, 300], dtype=dtype),
expected=np.array([0, 0, 0.5, 1, 1], dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.sinh,
np.array([1, 2, 3, 4], dtype=dtype),
expected=np.array(
[1.17520119, 3.62686041, 10.01787493, 27.2899172],
dtype=dtype))
self._assertOpOutputMatchesExpected(math_ops.sqrt,
np.array([[4, 9]],
dtype=dtype),
expected=np.array([[2, 3]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.tan,
np.array([1, 2, 3, 4], dtype=dtype),
expected=np.array(
[1.55740772, -2.18503986, -0.14254654, 1.15782128],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.tanh,
np.array([[1, 1, 1, 1], [1, 2, 3, 4]], dtype=dtype),
expected=np.array(
[[0.76159418, 0.76159418, 0.76159418, 0.76159418],
[0.76159418, 0.96402758, 0.99505478, 0.99932933]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
nn_ops.log_softmax,
np.array([[1, 1, 1, 1], [1, 2, 3, 4]], dtype=dtype),
expected=np.array(
[[-1.3862944, -1.3862944, -1.3862944, -1.3862944],
[-3.4401896, -2.4401896, -1.4401897, -0.44018969]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
nn_ops.elu,
np.array([[-1, 0, 1, -1e-6]], dtype=dtype),
expected=np.array([[-0.63212056, 0, 1, -9.999995e-07]],
dtype=dtype),
rtol=1e-5,
atol=1e-6)
self._assertOpOutputMatchesExpected(
nn_ops.selu,
np.array([[-1, 0, 1, -1e-5]], dtype=dtype),
expected=np.array(
[[-1.11133074, 0., 1.05070099, -1.758090550379974e-05]],
dtype=dtype),
rtol=1e-5,
atol=1e-6)
self._assertOpOutputMatchesExpected(nn_ops.relu,
np.array([[-1, 1]],
dtype=dtype),
expected=np.array([[0, 1]],
dtype=dtype))
self._assertOpOutputMatchesExpected(nn_ops.relu6,
np.array([[-0.05, 6.05, 5]],
dtype=dtype),
expected=np.array([[0, 6, 5]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
nn_ops.softmax,
np.array([[1, 1, 1, 1], [1, 2, 3, 4]], dtype=dtype),
expected=np.array(
[[0.25, 0.25, 0.25, 0.25],
[0.032058604, 0.087144323, 0.23688284, 0.64391428]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
nn_ops.softsign,
np.array([[-2, -1, 0, 1, 2]], dtype=dtype),
expected=np.array([[-0.66666669, -0.5, 0, 0.5, 0.66666669]],
dtype=dtype))
self._assertOpOutputMatchesExpected(
math_ops.is_finite,
np.array([[42, float("inf"), -123], [float("nan"), 0, -0.0]],
dtype=dtype),
expected=np.array([[True, False, True], [False, True, True]],
dtype=np.bool))
self._assertOpOutputMatchesExpected(
lambda x: array_ops.quantize_and_dequantize_v2(
x, -127, 127, True, 8),
np.array([-1, -0.5, 0, 0.3], dtype=dtype),
expected=np.array([-1, -64.0 / 127, 0, 38.0 / 127],
dtype=dtype))
