def forward()

in lib/normalize_ewma.py [0:0]

• 16 lines of code
• 3 McCabe index (conditional complexity)

    def forward(self, input_vector):
        # Make sure input is float32
        input_vector = input_vector.to(torch.float)

        if self.training:
            # Detach input before adding it to running means to avoid backpropping through it on
            # subsequent batches.
            detached_input = input_vector.detach()
            batch_mean = detached_input.mean(dim=tuple(range(self.norm_axes)))
            batch_sq_mean = (detached_input ** 2).mean(dim=tuple(range(self.norm_axes)))

            if self.per_element_update:
                batch_size = np.prod(detached_input.size()[: self.norm_axes])
                weight = self.beta ** batch_size
            else:
                weight = self.beta

            self.running_mean.mul_(weight).add_(batch_mean * (1.0 - weight))
            self.running_mean_sq.mul_(weight).add_(batch_sq_mean * (1.0 - weight))
            self.debiasing_term.mul_(weight).add_(1.0 * (1.0 - weight))

        mean, var = self.running_mean_var()
        return (input_vector - mean[(None,) * self.norm_axes]) / torch.sqrt(var)[(None,) * self.norm_axes]