Cmpatible loss function

[bva1986]

Hello

To reproduce results of AutoVC we used following loss function implementation on PyTorch

num_steps = 100000
G = Generator(32,256,512,32).to(device)
...
criterion_recon = nn.MSELoss().to(device)
criterion_recon0 = nn.MSELoss().to(device)
criterion_content = nn.L1Loss().to(device)
optimizer = optim.Adam(G.parameters(), lr=1e-4)
G.train()
...
for i in range(1, num_steps + 1):
    ...
    optimizer.zero_grad()
    X1rt, X1r, C1 = G(X1, S1, S1)
    #with torch.no_grad(): # Should we calculate C1r with grad?
    #    C1r = G(X1r[:, 0, :, :], S1, None)
    C1r = G(X1r[:, 0, :, :], S1, None)
    L_recon = criterion_recon(X1r[:, 0, :, :], X1)
    L_recon0 = criterion_recon0(X1rt[:, 0, :, :], X1)
    L_content = criterion_content(C1r, C1)
    loss = L_recon + 1.0*L_recon0 + 1.0*L_content
    loss.backward()
    optimizer.step()

But we did not achieved comparable voice quality and loss value around 1e-3 according issue whats your final loss and final learning rate? One of supposed reason is that we used inapropriate loss function implementation. Is our loss function implementation compatible with that you used?

BR

[auspicious3000]

yes, your loss function looks right

[bva1986]

Should we backpropogate Encoder->Decoder->Encoder or just Encoder when minimize loss C1r-C1 ?

C1r = G(X1r[:, 0, :, :], S1, None)
L_content = criterion_content(C1r, C1)

or

C1r = G(X1r[:, 0, :, :].detach(), S1, None)
L_content = criterion_content(C1r, C1)

or even

C1r = G(X1r[:, 0, :, :].detach(), S1, None)
L_content = criterion_content(C1r, C1.detach())

[auspicious3000]

The 1st one

[bva1986]

According to AutoVC presentation on ICML-2019 loss function is: L = Lrecon + lambda Lcontent. But according to paper https://arxiv.org/abs/1905.05879 loss function is L = Lrecon + mu Lrecon0 + lambda*Lcontent. Which of them is more preferable?

[auspicious3000]

The 2nd one

[bva1986]

Ok, thanks