nnf_nll_loss - ignore_index about torch

[dfalbel]

Yeah should be -100 follwing the pytorch impl:

def nll_loss(input, target, weight=None, size_average=None, ignore_index=-100,
             reduce=None, reduction='mean'):

[jwijffels]

Yes. I tested that also but it said boom on my Windows machine

[jwijffels]

What I meant to say is that this crashes my session at the call of cpp_torch_namespace_nll_loss_self_Tensor_target_Tensor

library(torch)
m = nn_log_softmax(dim=1)
input = torch_randn(3, 5, requires_grad=TRUE)
target = torch_tensor(c(1L, 0L, 4L))
input = m(input)
output = nnf_nll_loss(input, target, ignore_index=-100L)
output

while it should be calling https://github.com/mlverse/torch/blob/master/src/lantern/lantern.h#L1649

[dfalbel]

ok, I'll take a look ASAP

[dfalbel]

This works for me if I do:

target = torch_tensor(c(1L, 0L, 4L), dtype = torch_long())

I could consider making torch_long() the default dtype when converting from R integers to torch tensors. We did something similar for R doubles that are converted to Tensors with dtype = torch_float(). What do you think?

[jwijffels]

Indeed, works with long instead of int. Don't know enough about the C API of lantern/libtorch to give advice. I don't mind specifyng that it is a long. Don't know currently if this impacts speed of anything.