"view size is not compatible with input tensor's size and stride"

[cw-tan]

Hi, thanks for the great LBFGS optimizer. This issue is to bring to your attention an error that appeared for my use case.
I was using the LBFGS optimizer as a general purpose optimizer, with it being initialized as
optimizer = LBFGSNew([tensor_1, tensor_2], lr=lr, history_size=6, max_iter=4, line_search_fn=True) for example.

I got the following error during an optimization step:
RuntimeError: view size is not compatible with input tensor's size and stride (at least one dimension spans across two contiguous subspaces). Use .reshape(...) instead.

This error did not appear when I did optimizations with respect to single tensors, for example:
optimizer = LBFGSNew([tensor_1], lr=lr, history_size=6, max_iter=4, line_search_fn=True)

Based on this Pytorch issue, I'm guessing that the way that the separate tensors are (combined?) and stored in memory is not contiguous.

I managed to avoid the error by changing all the lines that used .view(...) to .reshape(...) or .contiguous().view(...), with these lines being lines 87, 89 and 108:

pytorch/lbfgsnew.py

Line 87 in 6c229b9

view = p.grad.data.to_dense().view(-1)

pytorch/lbfgsnew.py

Line 89 in 6c229b9

view = p.grad.data.view(-1)

pytorch/lbfgsnew.py

Line 108 in 6c229b9

new_param1=p.data.clone().view(-1)

My optimization seemed to work out well after this change (I opted for the .contiguous().view(...) change), but I just wanted to check if these changes won't create other problems as I'm not sure about the motivation for using .view(...) over .reshape(...) in the first place. Thank you!

[SarodYatawatta]

Thanks, I will check if the fix is OK, and merge it.

[SarodYatawatta]

Fixed. thanks

[cw-tan]

Thanks!