Deprecate redundant log-scale transform classes. #12832
Conversation
|
Looks like this barely changed a few images, but overall 👍 on the change. Surprising these aren't used anywhere.... |
|
The tiny change in baseline images appears to come from floating point errors between |
| with np.errstate(divide="ignore", invalid="ignore"): | ||
| log = {np.e: np.log, 2: np.log2, 10: np.log10}.get(self.base) | ||
| if log: # If possible, do everything in a single call to Numpy. | ||
| out = log(a) |
There was a problem hiding this comment.
Is this the optimization you are referring to? I guess I think this is a good idea since it covers 99.9% of the cases, and would instead suggest changing the images or increasing the tolerance. But I'm not an expert on how expensive the un-optimized code is...
060b128 to
8747e87
Compare
|
Decided to just bump the tolerances, which were already set to 0.02 on some architectures anyways. |
467f536 to
7a90d25
Compare
7a90d25 to
2df7888
Compare
|
rebased |
2df7888 to
d19fd81
Compare
…832-on-v3.1.x Backport PR #12832 on branch v3.1.x (Deprecate redundant log-scale transform classes.)
PR Summary
Having a bunch of separate log Transform classes is mentioned as an optimization, but that optimization wasn't even there anymore. Restore it, while keeping everything in a single log Transform class.
PR Checklist