Conversation
|
Is there a way to pass custom optimizer? |
|
Yes, directly using scipys minimize interface. Check here, section called "Custom minimizers". |
|
I see, thanks! |
|
Np! Ill see about the errors. I wasnt sure how to not compute the gradient with the |
pymc3/tuning/starting.py
Outdated
| else: | ||
| norm_grad = np.linalg.norm(grad) | ||
| self.progress.set_description(self.desc.format(neg_value, norm_grad)) | ||
|
|
|
@bwengals Not computing the grad with logp_dlogp_function is not possible at the moment. Shouldn't be hard to add however, I can do that if you like. |
|
@aseyboldt Absolutely if you don't mind. I can also give it a whack, but would like to ask you a couple Q's first about We could work on that after this PR is merged as a separate project too. Depending on how tricky it may be. What do folks think? |
|
Hey Bill, do you think you can squeeze the changes mentions below into your PR? |
|
#2523 is the PR where I implemented the "output both" feature in master. I agree that a kwarg is even better. |
|
Oh, so instead of output both, only output the human-readable version of the parameter values (with a kwarg option to include the |
|
I'd also like the kwarg-option. |
|
LGTM. Is it ready? |
|
If everyone's ok with it, I think so! |
|
Thanks bill! |
* small typo fix in ValueGradFunction docstring, extra_args -> extra_vars * added different interface to scipy optimizers, removed errors * change error if optimization result is bad to warning * remove test for optimization error, remove transform=None, update find_MAP call args * update find_MAP call arg * small docstring change * remove blank links, remove extraneous callback arg * remove unused imports * test fail with precision a bit off, 9.996 vs. 10. Switching to previous default method of BFGS. * removed optimization check (since 'return_raw' is an option), added 'include_transformed' * remove unused import * need include_transformed=True
6 participants
Fix for issue #2466. Closing #2468 in favor of this one.
method="L-BFGS-B"instead offmin=optimize.fmin_l_bfgs_blogp = inf, etc.) last good optimizer value is returned. If this happens, a warning is printed, instead of an error.