Hi, I’m trying to use PyMC to find the optimal parameters that describe some observed data, but it’s not working. I have two vectored parameters: X(x1,x2,xn) and V(v1,v2,…vn). I also have a pre-defined class that wraps a black-box function to take in X and V, returns simulated data d1, and it als…

You probably need to wrap your blackbox function with as_op. For example, see a similar discussion in Connecting PyMC3 to external code - help with understanding Theano custom Ops and more in https://discourse.pymc.io/search?q=as_op

@junpenglao thanks! But with as_op, will I still be able to use the gradient based samplers like HMCMC and NUTS?

I’d like to use a L2-norm likelihood function P(d|m) \propto exp \frac{||d-f(m)||^2}{\sigma^2} where d is the observed data, m are my parameters and f is the external function that returns simulated data. So can I just write an customized Theano function to return the difference between d and f(…

No, gradient information is not available using @as_op Yes, using a Normal likelihood with constant sigma (ie, not a latent variable) is the same as L2 regularization,

Ok, thanks! So in that case, should I write an extension of the theano.op class and place my gradient computation in the grad() method?

Hi, I implemented the tt.Op extension and it works with the metropolis sampler. However, I’m a bit stuck on getting the gradient to work (so I can use NUTS). It returns an error saying the object returned the wrong number of gradient terms. For my case, the inputs are 3 vector variables, so total…

Did you validate the gradient: http://docs.pymc.io/advanced_theano.html (see near the end)? Also, you might find recent PR from @mattpitkin helpful: https://github.com/pymc-devs/pymc3/pull/3169

I just tried to import the unit test tools but got an error: “No module named nose_paramterized”. Could this be a version issue? I’m running python 2.7 and pymc3 3.0, theano 0.9.0.dev.

Could you please try to upgrade your pymc3 and theano?

Using pm.DensityDist and customized likelihood with a black-box function

Questions

junpenglao August 21, 2018, 3:53pm 7

Yes. There are a few thread on that, see eg Custom theano Op to do numerical integration

Topic		Replies	Views
Back-box function without specifying gradient Questions	0	515	February 23, 2021
Custom likelihood from a python function v5 modeling	5	1102	May 27, 2024
Suggested workaround for blackbox likelihood example? Questions	6	819	February 8, 2021
How to calibrate the sigma of the likelihood within the black-box method? Questions	13	1555	June 30, 2019
Regarding likelihood evaluation in case of a blackbox likelihood Questions theano	4	816	October 9, 2020

Using pm.DensityDist and customized likelihood with a black-box function

Related topics