def forward()

in fast_grad_models.py [0:0]

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

    def forward(self, x):
        """
        Forward pass that returns also returns
        * the activations (H) and
        * the linear combinations (Z)
        of each layer, to be able to use the trick from [1].

        Args:
        - x : The inputs of the network
        Returns:
        - logits
        - activations at each layer (including the inputs)
        - linear combinations at each layer

        > [1] EFFICIENT PER-EXAMPLE GRADIENT COMPUTATIONS
        > by Ian Goodfellow
        > https://arxiv.org/pdf/1510.01799.pdf
        """
        x = x.view(-1, self.input_size)
        out = x

        # Save the model inputs, which are considered the activations of the 0'th layer.
        activations = [out]
        linearCombs = []

        for layer in self.hidden_layers:
            linearComb = layer(out)
            out = self.act(linearComb)

            # Save the activations and linear combinations from this layer.
            activations.append(out)
            linearComb.requires_grad_(True)
            linearComb.retain_grad()
            linearCombs.append(linearComb)

        logits = self.output_layer(out)

        logits.requires_grad_(True)
        logits.retain_grad()
        linearCombs.append(logits)

        return (logits, activations, linearCombs)