Use log1p(x) instead of log(1+x)

[kit1980]

Description

This function is more accurate than torch.log() for small values of input - https://pytorch.org/docs/stable/generated/torch.log1p.html

Found with TorchFix https://github.com/pytorch-labs/torchfix/

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactor

[xrennvidia]

Hi @kit1980

Thanks for catching this. log1p is more accurate for small values, and it should be same as log for big values? I just wonder if this is only a fix for small values, and how this will impact big values? Thanks.

[kit1980]

@xrennvidia I think it's the same for big values, only for small values there are numerical instabilities that log1p designed to address.

[xrennvidia]

@xrennvidia I think it's the same for big values, only for small values there are numerical instabilities that log1p designed to address.

Awesome, thanks a lot. I will approve the PR.

[xrennvidia]

LGTM.

[xrennvidia]

/te-ci pytorch L1