def forward(self, features_mean, features_variance):
features_stddev = torch.sqrt(features_variance)
div = features_mean / features_stddev
pdf = normpdf(div)
cdf = normcdf(div)
outputs_mean = features_mean * cdf + features_stddev * pdf
outputs_variance = (features_mean ** 2 + features_variance) * cdf \
+ features_mean * features_stddev * pdf - outputs_mean ** 2
if self._keep_variance_fn is not None:
outputs_variance = self._keep_variance_fn(outputs_variance)
return outputs_mean, outputs_variance
def _max_pool_internal(self, mu_a, mu_b, var_a, var_b):
stddev = torch.sqrt(var_a + var_b)
ab = mu_a - mu_b
alpha = ab / stddev
pdf = normpdf(alpha)
cdf = normcdf(alpha)
z_mu = stddev * pdf + ab * cdf + mu_b
z_var = ((mu_a + mu_b) * stddev * pdf + (mu_a**2 + var_a) * cdf +
(mu_b**2 + var_b) * (1.0 - cdf) - z_mu**2)
if self._keep_variance_fn is not None:
z_var = self._keep_variance_fn(z_var)
return z_mu, z_var
def forward(self, features_mean, features_variance):
features_stddev = torch.sqrt(features_variance)
div = features_mean / features_stddev
pdf = normpdf(div)
cdf = normcdf(div)
negative_cdf = 1.0 - cdf
mu_cdf = features_mean * cdf
stddev_pdf = features_stddev * pdf
squared_mean_variance = features_mean**2 + features_variance
mean_stddev_pdf = features_mean * stddev_pdf
mean_r = mu_cdf + stddev_pdf
variance_r = squared_mean_variance * cdf + mean_stddev_pdf - mean_r**2
mean_n = -features_mean * negative_cdf + stddev_pdf
variance_n = squared_mean_variance * negative_cdf - mean_stddev_pdf - mean_n**2
covxy = -mean_r * mean_n
outputs_mean = mean_r - self._negative_slope * mean_n
outputs_variance = variance_r \
+ self._negative_slope * self._negative_slope * variance_n \
- 2.0 * self._negative_slope * covxy
if self._keep_variance_fn is not None:
outputs_variance = self._keep_variance_fn(outputs_variance)
return outputs_mean, outputs_variance
