ENH: Adds pytorch 0.4 support #185
Conversation
thomasjpfan
force-pushed
the
pytorch-0.4
branch
from
4c23bcb
to
9233afa
Compare
ENH: Adds CyclicLR
Support (wildcard) param groups in initialization
| # currently, I see no solution that would result in a 1-dim | ||
| # LongTensor after passing through torch's DataLoader. If | ||
| # there is, we should use that instead. Otherwise, this will | ||
| # be obsolete once pytorch scalars arrive. |
There was a problem hiding this comment.
This comment is still valid though. This is a workaround for the case where you pass a tensor of shape (n,) as y to the data loader which will result in y.shape = (n,1) which we correct to its original shape (n,) here.
You can verify for yourself that this is still the case:
X = torch.zeros(10, 5)
y = torch.zeros(10)
d = TensorDataset(X, y)
l = DataLoader(d)
Xi, yi = next(iter(l))
assert len(yi.size()) == 1 # fails, is 2
There was a problem hiding this comment.
len(yi.size()) is 1 in pytorch 0.3.1 and 0.4.
There was a problem hiding this comment.
True, the problem lies with our to_tensor. We'll see what we can do.
…nto thomasjpfan-pytorch-0.4
This is because pytorch 0.4 no longer differentiates between Tensors and Variables.
translation example is fixed but still has issues, therefore WIP.
|
@BenjaminBossan I think this PR is ready. Do you think we can merge this so at least the master branch works with pytorch 0.4.0 until we release the next version? |
|
I can't merge before Wednesday ;) There are some references to Variables still left that could be removed, as well as a typo in your deprecation warning message, but that can be fixed later. |
|
@ottonemo I fixed the points I addressed earlier. If it's fine by you, I would merge now. |
Closes #184