Error while using dims with multiple linear regression model

1

Howdy! I ran into some problems using coords and dims in my model.
“wavelength” is 131, “compound” is 10, “samples” is 25 in the following code.

coords = {"wavelength": spectra.columns,
          "compound": conc.columns,
          "samples": spectra.index
          }

with pm.Model(coords=coords) as linearmodel:
  #data
  y = spectra.T.values
  x = conc.T.values
  #prior
  sigma = pm.HalfCauchy("sigma", beta=5)
  intercept = pm.Normal("intercept", 0, sigma=5, dims="wavelength")
  slope = pm.Normal("slope", 0, sigma = 5, dims=("wavelength", "compound"))

  #likelihood
  likelihood = pm.Normal("y", mu= (slope @ x) + intercept,sigma=sigma, observed=y)

Im getting the following error code
"ValueError: Incompatible Elemwise input shapes [(131, 25), (1, 131)]" for the sum of slope @ x and the intercept. If I replace dims=“wavelength” in the intercept for “shape=(131,1)” the code works as intended.
What would be the correct way to make this work using dims? I was under the impression that dims=“wavelength” or dims=("wavelength,) should make a (131,1) vector that would broadcast correctly.

2

dims in PyMC are just labels for the outputs of sampling, they have no effect. All broadcasting works positionally just like in numpy.

3

Ok! so there is no way to make the intercept a (131,1) vector by passing dims=“wavelength”.
What would be the recommended way to make a row or column vector with dims? adding an axis afterwards?

Like:

intercept = pm.Normal(0,1,dims="wavelength")

likelihood = pmNormal(mu= slope@x + intercept[None,:], sigma=sigma)
4

This is how I usually do it. You could also use .reshape. You could also transpose slope @ x, because from the error it seems the problem is that you have wavelengths on different dimensions in each object.