Mingbang Wang
Mingbang Wang
@frozenbugs @Rhett-Ying @peizhou001 @mfbalin @BarclayII I'd be grateful for your opinions and suggestions on this issue. Because this involves modifying the code written many years ago, I'm afraid I may...
> I think it's a historical issue that only `IdType` is used only that does not take more scenarios into consideration. > > Dynamically determining the type of `indptr`, `indices`,...
After further investigation, I've discovered a possible reason why we need `IdType`. It's because the approaches of cuda implementation of `COOToCSR` are different for `int32` and `int64`. See https://github.com/dmlc/dgl/blob/ed50c170dda9627730cb8ee4c7110205b6ea09de/src/array/cuda/coo2csr.cu#L25 and...
@mfbalin @frozenbugs See benchmark results in the description. The new implementation does not seem to be as efficient as we thought. Maybe we should keep it as is?
Hello @Ying-1106, it'd be helpful if you can provide the error message. And you can try `print(train_set)` to examine the training set and check if data is correct.