For example, the following code would silently corrupt the memory, but now it
generates an error:

A = torch.zeros(4, 4).cuda()
B = torch.zeros(3, 5).cuda()
idxs = torch.LongTensor([2, 1, 0]).cuda()
A.index_add_(0, idxs, B)  # or index_copy_
# RuntimeError: invalid argument 2: Source/destination tensor have different
# slice sizes (4 vs 5) at (...)/aten/src/THC/generic/THCTensorIndex.cu:46

Also, the following code generates a warning. (I'm not sure if anybody is
intentionally using this behavior, but I decided to play it safe.)

C = torch.zeros(3, 2, 2).cuda()
A.index_add_(0, idxs, C)  # or index_copy_
# Warning: source/destination slices have same size but different shape for
# an index operation.  This behavior is deprecated.

Hi, did anyone have a chance to look at this? Thanks!

@pytorchbot add to whitelist

thanks a lot @yongjik !

I think there might be something wrong with the index_copy_ function in the 0.4.0 release.
These are the variables and my goal is to copy the 1st and 2nd columns of B to the 1st and 2nd columns of A.

A = torch.zeros((3,5))
B = torch.zeros((3,2))
idxs = torch.LongTensor([0, 1])

When I use the following command, it will generate error message. (Dimension mismatching.)

A.index_copy_(1,idxs,B)

But when I transpose A and B and change the dim argument from 1 to 0, then it works.

torch.transpose(A,1,0).index_copy_(0,idxs,torch.transpose(B,1,0))