Error in beamsearch.py

[sahilm1992]

Hi,

I am getting the following error on line 96 in file Utils/beamsearch.py
On this line
self.mask = self.mask.gather(1, perm_mask)

Error is

gather_out_gpu(): expected dtype int64 for index

The issue is that perm_mask is having float values but torch.gather expects integer in the index argument

Could you please help me on this.

I used the same libraries with the version numbers recommended in the readme.

I am running for tsp10.

[chaitjo]

Hi @sahilm1992, thanks for your interest! That's unexpected, but maybe you can fix it by simply converting the float tensor to a long tensor?

E.g. self.mask = self.mask.gather(1, perm_mask.long())

Let me know if that works.

[sahilm1992]

let me try that. Could you confirm if the implementation works on gpu?

[chaitjo]

Yes, all experiments in the associated paper were run using this codebase. You can read about our exact hardware setup there. However, the code has not been updated since ~ 1 year. You can try the updated repository associated with the more comprehensive paper, too.

[chaitjo]

By the way, can you tell me your hardware setup? E.g. Operating system, Cuda version, GPU type, etc.

[sahilm1992]

It worked now.
Possibly some cuda issue.

Thanks :)

[chaitjo]

Great, feel free to open further issues if you need help.