batch_norm.py 文件源码

def get_output_for(self, input, deterministic=False, **kwargs):
        if deterministic:
            # use stored mean and std
            mean = self.mean
            std = self.std
        else:
            # use this batch's mean and std
            mean = input.mean(self.axes, keepdims=True)
            std = input.std(self.axes, keepdims=True)
            # and update the stored mean and std:
            # we create (memory-aliased) clones of the stored mean and std
            running_mean = theano.clone(self.mean, share_inputs=False)
            running_std = theano.clone(self.std, share_inputs=False)
            # set a default update for them
            running_mean.default_update = ((1 - self.alpha) * running_mean +
                                           self.alpha * mean)
            running_std.default_update = ((1 - self.alpha) * running_std +
                                          self.alpha * std)
            # and include them in the graph so their default updates will be
            # applied (although the expressions will be optimized away later)
            mean += 0 * running_mean
            std += 0 * running_std
        std += self.epsilon
        mean = T.addbroadcast(mean, *self.axes)
        std = T.addbroadcast(std, *self.axes)
        beta = T.addbroadcast(self.beta, *self.axes)
        gamma = T.addbroadcast(self.gamma, *self.axes)
        normalized = (input - mean) * (gamma / std) + beta
        return self.nonlinearity(normalized)