I’m analyzing time series data with a physics model and one of the parameters is a 2D vector, and I have priors on each of the two components, so I set up the model to sample them independently, and then create the vector from those: import numpy as np import pymc3 as pm with pm.Model(): a0 = p…

You can define a in single line to avoid the concatenation. Then, use use pm.Deterministic to define a function of RVs. Something like this (untested): from theano import tensor as T with pm.Model(): a = pm.Normal('a', mu=0, sd=1, shape=2) unit_vector = pm.Deterministic('unit_vector', a/T.norm…

I think T.sqrt(T.dot(a,a)) is a working denominator there. I’m finding that I’ll probably have to refactor the rest of my code to handle this anyhow, but this helps get me there, so thank you.

I often find that unit-normed parameters are hard to sample because the norm of the input vector is unconstrained. In this specific case, you may want to parametrize with a von-Mises distribution theta = pm.VonMises('theta') unit_vector = pm.Deterministic('unit_vector', T.cos(theta), T.sin(theta))

How to compute unit vector based on RVs?

Questions

junpenglao May 25, 2019, 5:16am 5

+1 to @chartl’s comment. You might also find this discussion useful: How to model sinusoids in white gaussian noise

Topic		Replies	Views
Unit vector for parameter values	2	50	September 29, 2024
pm.Potential and confliction between Theano and Numpy v5 theano , gaussian_process	9	669	March 2, 2023
Issue when trying to normalize random variable v5 theano , modeling , sampling	3	87	September 18, 2024
Cannot convert <class 'numpy.ndarray'> to TensorType' Questions	5	4067	February 13, 2020
Can't vectorise custom distribution Questions theano	2	759	May 24, 2018

How to compute unit vector based on RVs?

Related topics