Something is wrong in draw values
Do you reproduce the error with non-advi trace?
i had the same error while i was trying to run the code from the original ipynb https://github.com/pymc-devs/pymc3/blob/master/docs/source/notebooks/bayesian_neural_network_opvi-advi.ipynb Yes i had the same error with NUTS and Metropolis
I remember some changes there
We’ll try to solve the problem before release
If it is possible, could you please provide minimal failing example?
hi @rlouf , welcome! I think a great starting point is to run the notebooks in https://github.com/pymc-devs/pymc3/tree/master/docs/source/notebooks and try to fix/report the errors - we always need hands to maintain the notebooks as we are doing it mostly by hand currently. And our API changes quite a bit in the pass few months so something might be broken and we have not realized yet.
I think this is a great way also to learn python and bayesian statistics at the same time.
There are also notebooks need some explanations and background information, for example https://github.com/pymc-devs/pymc3/blob/master/docs/source/notebooks/cox_model.ipynb which would be a great starting point for contribution as well.
Hi @junpenglao ! Thank you for the pointers, it indeed looks like an interesting way to get to know PyMC3 and learn a bit more about bayesian statistics. I'll come back when my homework is done :) I might also add a comprehensive notebook on bayesian A/B testing (with hierarchical models and all)--unless you feel there is no need for it.
We just set up our discourse forum, for longer discussion please move to https://discourse.pymc.io/
Hi all! I've been using the ADVI in PyMC3 to fit a Poisson latent Gaussian model with ARD. In testing on simulated data, I've gotten good results with the old ADVI interface (in that the number of simulated relevant components is correctly recovered), but switching over to the new ADVI interface sometimes gives me inconsistent results. Should I expect these results to be roughly equivalent if I set
obj_optimizer=pm.adagrad_window(learning_rate), or are there other parameters I need to set appropriately?
I would recommend checking loss convergence first. Did it converge?
Great! I see you have already done a PR
Thanks Bill. I did that, I thought I had everything setup correctly but when I pass in x as observed I get an error. A shape mismatch error, input[0].shape[1] = Num records, but input[1].shape[1] = p. That's why I was looking for an example to study... if I have the model setup correctly I should be able to pass the 2d ndarray directly to observed right? I was thinking maybe it was a theano issue that.
My example works for the 1d case but fails for p>1. So I'm wondering if I need to wrap the 2d x in a theano object maybe?
try this:
gp = pm.gp.GP("gp", cov_func=cov, X=X, sigma=sn, observed=y)
setting X as an argument
also, were trying to move questions like this over to https://discourse.pymc.io/ . You'll probably get a quicker response there next time :)
hello, I have a hierarchical model which was working fine in PyMC 3.0 and, when I upgraded to 3.1 today, I get the
FloatingPointError: NaN occurred in optimization. error on the first ADVI iteration. All the parameters I set for the model (priors, etc.) have not changed, I've only changed the API to access ADVI to pm.fit. Am I missing something obvious with the upgrade?
@darribas you might find the discussion on discourse https://discourse.pymc.io/t/nan-occurred-in-optimization-with-nuts/39 and https://discourse.pymc.io/t/get-nan-or-inf-from-model-logp-model-test-point-is-an-attestation-of-incorrectly-configured-model/111 helpful.
Short answer is there might be some hidden problem in your initial model, if the above posts did not solve your problem, please open a discussion on discourse with your code and (simulated) data
oh is discourse being used now?
hi @hanzy1110 please use https://discourse.pymc.io/ to ask this type of questions as that is the official channel, we no longer use gitter.