fix safe_log on systems which use 32 bit floats #915
Conversation
src/iminuit/cost.py
Outdated
| @@ -116,7 +116,7 @@ class documentation for details. | |||
|
|
|||
| def _safe_log(x): | |||
| # guard against x = 0 | |||
| return np.log(x + 1e-323) | |||
| return np.log(np.maximum(np.finfo(x.dtype).smallest_normal, x)) | |||
There was a problem hiding this comment.
It's probably faster if np.finfo(x.dtype).smallest_normal is placed in a separate variable so that it doesn't need to be evaluated every time. You can find my current patch here (includes a pytest adjustment).
There was a problem hiding this comment.
In that case, we cannot select the subnormal that matches the dtype of x, but the current code does not work with numba anyway, so probably that's the solution to go for.
There was a problem hiding this comment.
I went for your solution, could you check again? Now Numba works again.
Unfortunately, it does not. You can find the logs for different architectures here. I do have access to these architectures in theory. I can try to look into it again and maybe test it with some verbose output to see what's going on there. |
|
Really nice to see that Debian is distributing iminuit. When I was younger, I used Debian for many years. |
|
I am going to merge this, because it is definitely more safe as the old code, since it can handle negative x (which should never happen, but you never know). If the issue persists, please reopen the issue. |
Closes #913 see that issue for a detailed analysis by @stephanlachnit why the original naive code did not work everywhere.
Maybe this is related to #902, too.
@stephanlachnit Could you please check whether this fix indeed solves the issue? I cannot run tests on the affected systems.