def forward()

in src/hyperconv.py [0:0]

• 21 lines of code
• 4 McCabe index (conditional complexity)

    def forward(self, x, z):
        '''
        :param x: the input signal as a B x ch_in x T tensor
        :param z: the weight-generating input as a B x z_dim x K tensor (K s.t. T is a multiple of K)
        :return: a B x ch_out x T tensor as the result of the hyper-convolution
        '''
        B = x.shape[0]
        assert x.shape[-1] % z.shape[-1] == 0
        # padding
        padding = self.dilation * (self.kernel_size - 1)
        x = F.pad(x, [padding, 0])
        # linearize input by appending receptive field in channels
        start, end = padding, x.shape[-1]
        x = th.cat([x[:, :, start-i*self.dilation:end-i*self.dilation] for i in range(self.kernel_size)], dim=1)
        # rearrange input to blocks for matrix multiplication
        x = x.permute(0, 2, 1).contiguous().view(x.shape[0] * z.shape[-1], x.shape[-1]//z.shape[-1], x.shape[1])
        # compute weights and bias
        weight = self.weight_model(z).view(B, self.ch_in * self.kernel_size, self.ch_out, z.shape[-1])
        weight = weight.permute(0, 3, 1, 2).contiguous().view(B * z.shape[-1], self.ch_in * self.kernel_size, self.ch_out)
        bias = self.bias_model(z).view(B, self.ch_out, z.shape[-1])
        bias = bias.permute(0, 2, 1).contiguous().view(B * z.shape[-1], self.ch_out)
        # compute result of dynamic convolution
        y = th.bmm(x, weight)
        y = y + bias[:, None, :]
        y = y.view(B, -1, self.ch_out).permute(0, 2, 1).contiguous()
        return y