def _batchnorm_scaling_terms_only(self, inp):
return B.batch_normalization(inputs=inp,
gamma=self.reshaped_gamma,
beta=0.0,
mean=0.0,
std=self.reshaped_std,
epsilon=self.epsilon)
def _build_activation_vars(self, input_act_vars):
new_shape = [(1 if (i != self.axis\
and i != (len(self._shape)+self.axis)) #neg self.axis
else self._shape[i])
for i in range(len(self._shape))]
self.reshaped_mean = self.mean.reshape(new_shape)
self.reshaped_std = self.std.reshape(new_shape)
self.reshaped_gamma = self.gamma.reshape(new_shape)
self.reshaped_beta = self.beta.reshape(new_shape)
return B.batch_normalization(inputs=input_act_vars,
gamma=self.reshaped_gamma,
beta=self.reshaped_beta,
mean=self.reshaped_mean,
std=self.reshaped_std,
epsilon=self.epsilon)