Indexing embedding using longtensor loses gradient_fn

1

When I use a long tensor to index embedding, it loses its grad_fn:

embed_option = nn.Embedding(4, 10)
vec = embed_option(torch.LongTensor([[0,1,2]]))

vec.requires_grad is Flase.

When I tried in main scope, everything worked fine. However, if I define this in a nn.Module class, it loses its gradient function. Any idea how this could happen?

2

fk me. I wrapped it in with torch.no_grad() in caller function.