Inconsistent/Confusing behavior of util "safe_div"

[aobo-y]

https://github.com/pytorch/captum/blob/master/captum/_utils/common.py#L26-L37

The behavior of how safe_div uses default_value is inconsistent depending on if denorm is float or tensor

Described in the the comments, seems the default_value should be used directly as the devision result when denorm is 0. This is indeed implemented accordingly when denorm is float

return (numerator / denom) if denom != 0.0 else default_value

However, when denorm is tensor, the default_value is actually used as the default denorm instead

return numerator / torch.where(denom != 0.0, denom, default_value)

I think the 2nd behavior is more reasonable, i.e., when denorm is float, update to

return numerator / (denom if denom != 0.0 else default_value)

Also better rename the parameter to default_denorm to avoid further confusion.
The question left is if there is any need or use cases for the 1st bahvior.

[NarineK]

@aobo-y, that's a good point! I agree, let's change it to return numerator / (denom if denom != 0.0 else default_value).
default_denorm is a good name.
Thank you!

[aobo-y]

fixed in #751