loss nan

[rangmiao]

hello, i have a problem. when I use ffhq data with 256x256, g loss and d loss is nan.

here is log:
D_loss: nan, D_loss_grad_norm: nan, D_lr: 0.001882
D_reg: 0.002352, D_reg_grad_norm: 0.001569, G_loss: nan
G_loss_grad_norm: nan, G_lr: 0.0016, G_reg: nan
G_reg_grad_norm: nan, pl_avg: nan, seen: 15
0%| | 14/1000000 [00:51<634:57:52, 2.29s/it]

[adriansahlman]

Hey,

Is the loss nan from the very first iteration? Or does it become nan after a couple of iterations?

What settings do you use when you run the training?