The derivative for torch.bernoulli() is removed in 0.4.1? 0.4.0 still has it

1

I got an error when do backward()

“the derivative for _th_bernoulli is not implemented”

2

you can’t backprop through sampling functions (unless they are reparametrized). Even in 0.4, it probably returned all zero gradient and that was a bug if it was the case.