def forward()

in janus/models/vq_model.py [0:0]

• 35 lines of code
• 5 McCabe index (conditional complexity)

    def forward(self, z):
        # reshape z -> (batch, height, width, channel) and flatten
        z = torch.einsum("b c h w -> b h w c", z).contiguous()
        z_flattened = z.view(-1, self.e_dim)
        # distances from z to embeddings e_j (z - e)^2 = z^2 + e^2 - 2 e * z

        if self.l2_norm:
            z = F.normalize(z, p=2, dim=-1)
            z_flattened = F.normalize(z_flattened, p=2, dim=-1)
            embedding = F.normalize(self.embedding.weight, p=2, dim=-1)
        else:
            embedding = self.embedding.weight

        d = (
            torch.sum(z_flattened**2, dim=1, keepdim=True)
            + torch.sum(embedding**2, dim=1)
            - 2
            * torch.einsum(
                "bd,dn->bn", z_flattened, torch.einsum("n d -> d n", embedding)
            )
        )

        min_encoding_indices = torch.argmin(d, dim=1)
        z_q = embedding[min_encoding_indices].view(z.shape)
        perplexity = None
        min_encodings = None
        vq_loss = None
        commit_loss = None
        entropy_loss = None

        # compute loss for embedding
        if self.training:
            vq_loss = torch.mean((z_q - z.detach()) ** 2)
            commit_loss = self.beta * torch.mean((z_q.detach() - z) ** 2)
            entropy_loss = self.entropy_loss_ratio * compute_entropy_loss(-d)

        # preserve gradients
        z_q = z + (z_q - z).detach()

        # reshape back to match original input shape
        z_q = torch.einsum("b h w c -> b c h w", z_q)

        return (
            z_q,
            (vq_loss, commit_loss, entropy_loss),
            (perplexity, min_encodings, min_encoding_indices),
        )