Multiple independent random walks with differing innovation sds

[martiningram]

Hi all,

Following on from a post on discourse, Junpeng asked me to open an issue.

I am trying to declare a GaussianRandomWalk prior on multiple time series, each independent and with its own random walk standard deviation. I tried:

mean_sigma = pm.HalfNormal('mean_rw_sigma', 1., shape=n_tourneys)
prior_means = pm.GaussianRandomWalk('prior_means', sigma=mean_sigma, shape=(n_years-1, n_tourneys))

but this does not compile, giving the error:

ValueError: Input dimension mis-match. (input[0].shape[1] = 3, input[1].shape[1] = 4)

Junpeng mentioned that this likely used to work but was broken by a helper. He also mentioned a workaround with a diagonal covariance matrix, which should work but presumably this version would be more efficient.

Versions and main components

• PyMC3 Version: 3.10.0
• Theano Version: 1.0.14
• Python Version: 3.7.9
• Operating system: Ubuntu
• How did you install PyMC3: pip

[junpenglao]

To add some context, the helper function is assuming the first dimension is time dimension. One workaround is to check the shape and only execute if the mu.shape[0] == shape[0]

[Sayam753]

I think the issue is similar to #4010

[junpenglao]

You are right, mark as duplicate.