def testAffineSoftplusIsCentered(self):
"""Check that affine_softplus(0) == 1."""
for _ in range(10):
lo = np.random.uniform(0., 0.1)
ref = np.random.uniform(0.2, 10.)
y = util.affine_softplus(np.array(0.), lo=lo, ref=ref)
self.assertAllClose(y, ref)
def testAffineSoftplusSpansRange(self):
"""Check that affine_softplus()'s output is in [lo, infinity]."""
x = np.finfo(np.float32).max * np.array([-1, 1], dtype=np.float32)
for _ in range(10):
lo = np.random.uniform(0., 0.1)
ref = np.random.uniform(0.2, 10.)
y = util.affine_softplus(x, lo=lo, ref=ref)
self.assertAllClose(y[0], lo)
self.assertAllGreater(y[1], 1e10)
def testAffineSoftplusRoundTrip(self):
"""Check that x = inv_affine_softplus(affine_softplus(x)) in general."""
x = np.float32(np.linspace(-10., 10., 1000))
for _ in range(10):
lo = np.random.uniform(0., 0.1)
ref = np.random.uniform(0.2, 10.)
y = util.affine_softplus(x, lo=lo, ref=ref)
x_recon = util.inv_affine_softplus(y, lo=lo, ref=ref)
self.assertAllClose(x, x_recon, atol=1e-5, rtol=1e-3)
def _construct_scale(x, scale_lo, scale_init, float_dtype):
"""Helper function for constructing scale variables."""
if scale_lo == scale_init:
# If the difference between the minimum and initial scale is zero, then
# we just fix `scale` to be a constant.
scale = tf.tile(
tf.cast(scale_init, float_dtype)[tf.newaxis, tf.newaxis],
(1, x.shape[1]))
else:
# Otherwise we construct a "latent" scale variable and define `scale`
# As an affine function of a softplus on that latent variable.
latent_scale = tf.compat.v1.get_variable(
'LatentScale', initializer=tf.zeros((1, x.shape[1]), float_dtype))
scale = util.affine_softplus(latent_scale, lo=scale_lo, ref=scale_init)
return scale
def scale(self):
"""Returns the loss's current scale parameters.
Returns:
a TF tensor of size (1, self._num_channels) and type self._float_dtype,
containing the current estimated scale parameter for each channel,
which will presumably change during optimization. This tensor is a
function of the latent scale tensor being optimized over, and is not a
TF variable itself.
"""
if self._scale_lo == self._scale_init:
# If the difference between the minimum and initial scale is zero, then
# we just fix `scale` to be a constant.
return tf.tile(
tf.cast(self._scale_init, self._float_dtype)[tf.newaxis, tf.newaxis],
(1, self._num_channels))
else:
return util.affine_softplus(
self._latent_scale, lo=self._scale_lo, ref=self._scale_init)
def testDefaultAffineSoftplusRoundTrip(self):
"""Check that x = inv_affine_softplus(affine_softplus(x)) by default."""
x = np.float32(np.linspace(-10., 10., 1000))
y = util.affine_softplus(x)
x_recon = util.inv_affine_softplus(y)
self.assertAllClose(x, x_recon, atol=1e-5, rtol=1e-3)
