A clarification in the code-base

[KawshikManikantan]

In the file: src/model/mention_proposal/utils.py Line 9: sort_scores = ment_starts + 1e-5 * ment_starts Should it be + 1e-5 * ment_ends - To have shorter mentions (starting at the same index) first? Or is it just some coding convenience?

[shtoshni]

@KawshikManikantan Thanks for raising this issue. You're right, I have fixed it in the latest commit. Let me know if it looks alright.

[KawshikManikantan]

@shtoshni Thanks for the immediate reply. Although the above change should work. I still find some discrepancies. For example in the figure below. The first tensor is the ment_starts[sorted_indices] and the second is ment_ends[sorted_indices]. We can see that 306 306 pair in the first tensor has a correct 306 308 pair while it is not the case with 960 960 pair which has a corresponding 961 960. Can you think of any reason why this happens?

[KawshikManikantan]

@shtoshni It looks like an error due to floating point inaccuracies. This works as far as I can see: sort_scores = ment_starts.to(torch.float64) + 1e-5 * ment_ends.to(torch.float64)

[shtoshni]

@KawshikManikantan: Must be due to difference in torch version. I'm not running into any such issue. Regarding the sorting, yeah that looks weird tbh, no idea why that's the case.

[KawshikManikantan]

ment_starts = torch.tensor([960.0,960.0]) 
ment_ends = torch.tensor([961.0,960.0]) 
torch.set_printoptions(precision=10) 
sort_scores = ment_starts + 0.00001 * ment_ends 
print(sort_scores,sort_scores.dtype)

Output: tensor([960.0095825195, 960.0095825195]) torch.float32

ment_starts = torch.tensor([960.0,960.0],dtype=torch.float64)
ment_ends = torch.tensor([961.0,960.0],dtype=torch.float64)
torch.set_printoptions(precision=10)
sort_scores = ment_starts + 0.00001 * ment_ends
print(sort_scores,sort_scores.dtype)

Output: tensor([960.0096100000, 960.0096000000], dtype=torch.float64) torch.float64

This is the result I get in multiple environments and the reason for my sorting troubles. It resolves upon typecasting though.