Error in hybrid loss

[hummat]

The hybrid loss from eq. 5 is defined (and also accordingly implemented in the official implementation) as:

losses = vb_losses + self.hybrid_coeff * ce_losses

Here it is implemented as follows though:

d3pm/d3pm_runner.py

Line 294 in 3ceb637

return self.hybrid_loss_coeff * vb_loss + ce_loss, {

[ramithuh]

in that case, for the MNIST example, it seems like it's training only based on ce_loss, right?

d3pm/d3pm_runner.py

Lines 294 to 297 in 3ceb637

	return self.hybrid_loss_coeff * vb_loss + ce_loss, {
	"vb_loss": vb_loss.detach().item(),
	"ce_loss": ce_loss.detach().item(),
	}

d3pm/d3pm_runner.py

Line 345 in 3ceb637

d3pm = D3PM(DummyX0Model(1, N), 1000, num_classes=N, hybrid_loss_coeff=0.0).cuda()