Arviz.summary() ValueError

Questions v5
1

Hello,
I’m trying to use the arviz summary method like
az.summary(idata.prior)
but keep getting:
ValueError: could not broadcast input array from shape (9,3) into shape (9,)

My prior has the following dimensions:
(chain: 1, draw: 500, aoa: 297, spline: 40)

Has anyone else encountered the same issue with arviz or knows how to fix it?
Thanks in advance.

2

This should work (although it’s more “clean” to select the prior group with az.summary(idata, group='prior').

Could you share a bit more of your model? This doesn’t seem like an error coming from az.summary, I would expect it instead from a shape issue in the model code.

3

I’m working with a quite simple cubic spline model (using patsy) that looks like that:

knot_list = np.arange(0, 360, 10)
B = dmatrix(
    "bs(ALPHA, knots=knots, degree=3, include_intercept=True) - 1",
    {"ALPHA": df['ALPHA'], "knots": knot_list},
)

with pm.Model() as model:

    model.add_coord('spline', length=B.shape[1], mutable=True)
    model.add_coord('aoa', df['ALPHA'], mutable=True)

    data_b = pm.MutableData('data_b', B, dims=('aoa', 'spline'))

    tau = pm.HalfCauchy('tau', 20) 
    beta = pm.Normal("beta", mu=0, sigma=tau, dims='spline')
    mu = pm.Deterministic("mu", pm.math.dot(data_b, beta), dims='aoa')
    sigma = pm.HalfNormal("sigma", 0.01)
    c_predict = pm.Normal("c_predict", mu, sigma, observed=df['CZ'], dims='aoa')
4

I copy-pasted your code, adding:

    idata = pm.sample_prior_predictive()
az.summary(idata, group='prior', var_names=['beta', 'tau', 'sigma'])

and it ran fine. How are you generating your prior draws?

1 Like
5

Apparently it was about the var_names parameter as it ran fine now after adding it. Thanks a lot!

6

var_names just selects a subset of variables to display in the results table. I did that to keep it from being too long, but it also worked fine without it. Does it not for you?

7

No, without var_names it still doesn’t work. I’m on arviz 0.16 and I also used pm.sample_prior_predictive().

9

@jan5 I was running into a similar issue and just diagnosed it, but judging from your comments it might have had a different cause.

My issue was I was using the coords argument to subset some of my fixed-effects parameters, but the set of coordinates I was passing included repeated values. This gave me a similar ValueError (though the shape it was trying to broadcast was (9, 2) rather than (9, 3) like you’re seeing).

Probably a different root cause as it seems like you weren’t using the coords argument, but sharing just in case it helps you diagnose your issue.

2 Likes