def _testSoftplus(self, np_features, use_gpu=False):
np_features = np.asarray(np_features)
np_softplus = self._npSoftplus(np_features)
with self.test_session(use_gpu=use_gpu) as sess:
softplus = nn_ops.softplus(np_features)
softplus_inverse = distribution_util.softplus_inverse(softplus)
[tf_softplus, tf_softplus_inverse] = sess.run([
softplus, softplus_inverse])
self.assertAllCloseAccordingToType(np_softplus, tf_softplus)
rtol = {"float16": 0.07, "float32": 0.003, "float64": 0.002}.get(
str(np_features.dtype), 1e-6)
# This will test that we correctly computed the inverse by verifying we
# recovered the original input.
self.assertAllCloseAccordingToType(
np_features, tf_softplus_inverse,
atol=0., rtol=rtol)
self.assertAllEqual(np.ones_like(tf_softplus).astype(np.bool),
tf_softplus > 0)
self.assertShapeEqual(np_softplus, softplus)
self.assertShapeEqual(np_softplus, softplus_inverse)
self.assertAllEqual(np.ones_like(tf_softplus).astype(np.bool),
np.isfinite(tf_softplus))
self.assertAllEqual(np.ones_like(tf_softplus_inverse).astype(np.bool),
np.isfinite(tf_softplus_inverse))
def _testSoftplus(self, np_features, use_gpu=False):
np_features = np.asarray(np_features)
np_softplus = self._npSoftplus(np_features)
with self.test_session(use_gpu=use_gpu) as sess:
softplus = nn_ops.softplus(np_features)
softplus_inverse = distribution_util.softplus_inverse(softplus)
[tf_softplus,
tf_softplus_inverse] = sess.run([softplus, softplus_inverse])
self.assertAllCloseAccordingToType(np_softplus, tf_softplus)
rtol = {
"float16": 0.07,
"float32": 0.003,
"float64": 0.002
}.get(str(np_features.dtype), 1e-6)
# This will test that we correctly computed the inverse by verifying we
# recovered the original input.
self.assertAllCloseAccordingToType(np_features,
tf_softplus_inverse,
atol=0.,
rtol=rtol)
self.assertAllEqual(
np.ones_like(tf_softplus).astype(np.bool), tf_softplus > 0)
self.assertShapeEqual(np_softplus, softplus)
self.assertShapeEqual(np_softplus, softplus_inverse)
self.assertAllEqual(
np.ones_like(tf_softplus).astype(np.bool),
np.isfinite(tf_softplus))
self.assertAllEqual(
np.ones_like(tf_softplus_inverse).astype(np.bool),
np.isfinite(tf_softplus_inverse))
def testInverseSoftplusGradientNeverNan(self):
with self.test_session():
# Note that this range contains both zero and inf.
x = constant_op.constant((10.**np.arange(-8, 6)).astype(np.float16))
y = distribution_util.softplus_inverse(x).eval()
# Equivalent to `assertAllFalse` (if it existed).
self.assertAllEqual(np.zeros_like(y).astype(np.bool), np.isnan(y))
def testInverseSoftplusGradientNeverNan(self):
with self.test_session():
# Note that this range contains both zero and inf.
x = constant_op.constant(np.logspace(-8, 6).astype(np.float16))
y = distribution_util.softplus_inverse(x)
grads = gradients_impl.gradients(y, x)[0].eval()
# Equivalent to `assertAllFalse` (if it existed).
self.assertAllEqual(np.zeros_like(grads).astype(np.bool), np.isnan(grads))
def testInverseSoftplusGradientNeverNan(self):
with self.test_session():
# Note that this range contains both zero and inf.
x = constant_op.constant((10.**np.arange(-8,
6)).astype(np.float16))
y = distribution_util.softplus_inverse(x).eval()
# Equivalent to `assertAllFalse` (if it existed).
self.assertAllEqual(np.zeros_like(y).astype(np.bool), np.isnan(y))
def testInverseSoftplusGradientNeverNan(self):
with self.test_session():
# Note that this range contains both zero and inf.
x = constant_op.constant(np.logspace(-8, 6).astype(np.float16))
y = distribution_util.softplus_inverse(x)
grads = gradients_impl.gradients(y, x)[0].eval()
# Equivalent to `assertAllFalse` (if it existed).
self.assertAllEqual(
np.zeros_like(grads).astype(np.bool), np.isnan(grads))
def _inverse_and_inverse_log_det_jacobian(self, y):
event_dims = self._event_dims_tensor(y)
# Could also do:
# ildj = math_ops.reduce_sum(y - distribution_util.softplus_inverse(y),
# axis=event_dims)
# but the following is more numerically stable. Ie,
# Y = Log[1 + exp{X}] ==> X = Log[exp{Y} - 1]
# ==> dX/dY = exp{Y} / (exp{Y} - 1)
# = 1 / (1 - exp{-Y}),
# which is the most stable for large Y > 0. For small Y, we use
# 1 - exp{-Y} approx Y.
ildj = -math_ops.reduce_sum(math_ops.log(-math_ops.expm1(-y)),
axis=event_dims)
return distribution_util.softplus_inverse(y), ildj
def _inverse(self, y): return distribution_util.softplus_inverse(y)
