experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Fixes unit tests for inverse hyperbolic functions that were failing because numeric gradients were computed too close to a branch cut (for complex arguments) or singularity (for real arguments) where the function is not differentiable (See, e.g., http://mathworld.wolfram.com/BranchCut.html). This change moves the test points away from the branch cut/singularity.

Browse Source 
Improves precision of double precision numerical gradients by using a smaller step size delta (the optimal for symmetric difference approximation with functions computed with O(epsilon) error is epsilon^(1/3), so for double64 it is ~1e-5).

PiperOrigin-RevId: 163706297
This commit is contained in:
A. Unique TensorFlower 2017-07-31 09:53:35 -07:00 committed by TensorFlower Gardener
parent 99b190a1f1
commit 29550762bd
1 changed files with 20 additions and 25 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

45
tensorflow/python/kernel_tests/cwise_ops_test.py
View File

@ -100,18 +100,18 @@ class UnaryOpTest(test.TestCase):
inxf = ops.convert_to_tensor(xf)
yf = tf_func(inxf)
_, jacob_n = gradient_checker.compute_gradient(
inxf, s, yf, s, x_init_value=xf)
inxf, s, yf, s, x_init_value=xf, delta=1e-2)
jacob_n = jacob_n.astype(np.float16)
self.assertAllClose(jacob_t, jacob_n, rtol=5e-3, atol=5e-3)
elif x.dtype in (np.float32, np.complex64):
s = list(np.shape(x))
jacob_t, jacob_n = gradient_checker.compute_gradient(
inx, s, y, s, x_init_value=x)
inx, s, y, s, x_init_value=x, delta=1e-3)
self.assertAllClose(jacob_t, jacob_n, rtol=1e-3, atol=1e-3)
elif x.dtype in (np.float64, np.complex128):
s = list(np.shape(x))
jacob_t, jacob_n = gradient_checker.compute_gradient(
inx, s, y, s, x_init_value=x)
inx, s, y, s, x_init_value=x, delta=1e-5)
self.assertAllClose(jacob_t, jacob_n, rtol=1e-5, atol=1e-5)
def _check(self, result_tensor, result_np, input_sp_t, tol):
@ -184,7 +184,7 @@ class UnaryOpTest(test.TestCase):
def testFloatBasic(self):
x = np.arange(-3, 3).reshape(1, 3, 2).astype(np.float32)
w = x - x.min() + 1.01 # all greater than 1
w = x - x.min() + 1.02 # all greater than 1
y = (x + .5).astype(np.float32) # no zero
z = (x + 15.5).astype(np.float32) # all positive
k = np.arange(-0.90, 0.90, 0.25).astype(np.float32) # between -1 and 1
@ -204,10 +204,9 @@ class UnaryOpTest(test.TestCase):
self._compareBoth(x, np.sinh, math_ops.sinh)
self._compareBoth(x, np.cosh, math_ops.cosh)
self._compareBoth(x, np.tanh, math_ops.tanh)
# b/63457572: failing under CUDA.
# self._compareBoth(x, np.arcsinh, math_ops.asinh)
# self._compareBoth(w, np.arccosh, math_ops.acosh)
# self._compareBoth(k, np.arctanh, math_ops.atanh)
self._compareBoth(x, np.arcsinh, math_ops.asinh)
self._compareBoth(w, np.arccosh, math_ops.acosh)
self._compareBoth(k, np.arctanh, math_ops.atanh)
self._compareBoth(x, self._sigmoid, math_ops.sigmoid)
self._compareBoth(x, self._log_sigmoid, math_ops.log_sigmoid)
self._compareBoth(y, np.sign, math_ops.sign)
@ -253,8 +252,7 @@ class UnaryOpTest(test.TestCase):
self._compareBoth(x, np.log, math_ops.log)
self._compareBoth(x, np.log1p, math_ops.log1p)
self._compareBoth(x, np.sinh, math_ops.sinh)
# b/63457572.
# self._compareBoth(x, np.arcsinh, math_ops.asinh)
self._compareBoth(x, np.arcsinh, math_ops.asinh)
self._compareBoth(x, np.cosh, math_ops.cosh)
self._compareBoth(x, np.tanh, math_ops.tanh)
self._compareBoth(x, self._sigmoid, math_ops.sigmoid)
@ -280,7 +278,7 @@ class UnaryOpTest(test.TestCase):
def testDoubleBasic(self):
x = np.arange(-3, 3).reshape(1, 3, 2).astype(np.float64)
w = x - x.min() + 1.01 # all greater than 1
w = x - x.min() + 1.02 # all greater than 1
y = (x + .5).astype(np.float64) # no zero
z = (x + 15.5).astype(np.float64) # all positive
k = np.arange(-0.90, 0.90, 0.35).reshape(1, 3, 2).astype(
@ -300,10 +298,9 @@ class UnaryOpTest(test.TestCase):
self._compareBoth(x, np.sinh, math_ops.sinh)
self._compareBoth(x, np.cosh, math_ops.cosh)
self._compareBoth(x, np.tanh, math_ops.tanh)
# b/63457572: failing under CUDA.
# self._compareBoth(x, np.arcsinh, math_ops.asinh)
# self._compareBoth(w, np.arccosh, math_ops.acosh)
# self._compareBoth(k, np.arctanh, math_ops.atanh)
self._compareBoth(x, np.arcsinh, math_ops.asinh)
self._compareBoth(w, np.arccosh, math_ops.acosh)
self._compareBoth(k, np.arctanh, math_ops.atanh)
self._compareBoth(x, self._sigmoid, math_ops.sigmoid)
self._compareBoth(y, np.sign, math_ops.sign)
self._compareBoth(x, np.sin, math_ops.sin)
@ -394,7 +391,7 @@ class UnaryOpTest(test.TestCase):
def testComplex64Basic(self):
x = np.complex(1, 1) * np.arange(-3, 3).reshape(1, 3,
2).astype(np.complex64)
y = x + 0.5 # no zeros
y = x + np.complex(0.5, 0.5) # no zeros
self._compareBoth(x, np.abs, math_ops.abs)
self._compareBoth(x, np.abs, _ABS)
self._compareBoth(x, np.negative, math_ops.negative)
@ -410,10 +407,9 @@ class UnaryOpTest(test.TestCase):
self._compareCpu(x, np.sinh, math_ops.sinh)
self._compareCpu(x, np.cosh, math_ops.cosh)
self._compareCpu(x, np.tanh, math_ops.tanh)
# b/63457572: failing under CUDA.
# self._compareCpu(x, np.arcsinh, math_ops.asinh)
# self._compareCpu(x, np.arccosh, math_ops.acosh)
# self._compareCpu(x, np.arctanh, math_ops.atanh)
self._compareCpu(y, np.arcsinh, math_ops.asinh)
self._compareCpu(y, np.arccosh, math_ops.acosh)
self._compareCpu(y, np.arctanh, math_ops.atanh)
self._compareCpu(x, self._sigmoid, math_ops.sigmoid)
self._compareCpu(x, np.sin, math_ops.sin)
self._compareCpu(x, np.cos, math_ops.cos)
@ -434,7 +430,7 @@ class UnaryOpTest(test.TestCase):
def testComplex128Basic(self):
x = np.complex(1, 1) * np.arange(-3, 3).reshape(1, 3,
2).astype(np.complex128)
y = x + 0.5 # no zeros
y = x + np.complex(0.5, 0.5) # no zeros
self._compareBoth(x, np.abs, math_ops.abs)
self._compareBoth(x, np.abs, _ABS)
self._compareBoth(x, np.negative, math_ops.negative)
@ -450,10 +446,9 @@ class UnaryOpTest(test.TestCase):
self._compareCpu(x, np.sinh, math_ops.sinh)
self._compareCpu(x, np.cosh, math_ops.cosh)
self._compareCpu(x, np.tanh, math_ops.tanh)
# b/63457572, failing under CUDA.
# self._compareCpu(x, np.arcsinh, math_ops.asinh)
# self._compareCpu(x, np.arccosh, math_ops.acosh)
# self._compareCpu(x, np.arctanh, math_ops.atanh)
self._compareCpu(y, np.arcsinh, math_ops.asinh)
self._compareCpu(y, np.arccosh, math_ops.acosh)
self._compareCpu(y, np.arctanh, math_ops.atanh)
self._compareCpu(x, self._sigmoid, math_ops.sigmoid)
self._compareCpu(x, np.sin, math_ops.sin)
self._compareCpu(x, np.cos, math_ops.cos)