weepiess

> When I set use affine coupling rather than additive coupling, training on cifar10, n_flow = 32, n_block = 3, the weights will rapidly become Nan, and sample s become...

I have the same problem.