Quickstart regression example fails -- ValueError: setting an array element with a sequence.

[ahwillia]

Just installed pymc3 and am having some trouble. Some simple examples seem to be working, but the example on the front page fails:

import pymc3 as pm
import numpy.random as npr
X = npr.randn(100, 1)
w = npr.randn(1)
y = X @ w + (npr.randn(100) * .1)

with pm.Model() as linear_model:
    weights = pm.Normal('weights', mu=0, sigma=1)
    noise = pm.Gamma('noise', alpha=2, beta=1)
    y_observed = pm.Normal('y_observed',
                mu=X.dot(weights),
                sigma=noise,
                observed=y)

    posterior = pm.sample()

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-6-f503c46c891e> in <module>
     11                 mu=X.dot(weights),
     12                 sigma=noise,
---> 13                 observed=y)
     14 
     15     posterior = pm.sample()

~/anaconda3/lib/python3.7/site-packages/pymc3/distributions/distribution.py in __new__(cls, name, *args, **kwargs)
     43                 raise TypeError("observed needs to be data but got: {}".format(type(data)))
     44             total_size = kwargs.pop('total_size', None)
---> 45             dist = cls.dist(*args, **kwargs)
     46             return model.Var(name, dist, data, total_size)
     47         else:

~/anaconda3/lib/python3.7/site-packages/pymc3/distributions/distribution.py in dist(cls, *args, **kwargs)
     54     def dist(cls, *args, **kwargs):
     55         dist = object.__new__(cls)
---> 56         dist.__init__(*args, **kwargs)
     57         return dist
     58 

~/anaconda3/lib/python3.7/site-packages/pymc3/distributions/continuous.py in __init__(self, mu, sigma, tau, sd, **kwargs)
    467         self.tau = tt.as_tensor_variable(tau)
    468 
--> 469         self.mean = self.median = self.mode = self.mu = mu = tt.as_tensor_variable(floatX(mu))
    470         self.variance = 1. / self.tau
    471 

~/anaconda3/lib/python3.7/site-packages/pymc3/theanof.py in floatX(X)
     63     """
     64     try:
---> 65         return X.astype(theano.config.floatX)
     66     except AttributeError:
     67         # Scalar passed

ValueError: setting an array element with a sequence.

Interestingly, the following simpler example does work:

with pm.Model() as linear_model:
    mu = pm.Normal('mu', mu=0, sigma=1)
    noise = pm.Gamma('noise', alpha=2, beta=1)
    y_observed = pm.Normal('y_observed',
                mu=mean,
                sigma=noise,
                observed=npr.randn(100))

    posterior = pm.sample()

pymc3 version: 3.7
theano version: 1.0.4
Python version: 3.7.1

Edit: I get the same error if I try different dimensions (e.g. w=npr.randn(100, 5))

[ColCarroll]

Thank you for reporting - this is an important problem to fix!

I can confirm I reproduce it on my machine, and here is a simple fix (swapping X.dot(weights) for pm.math.dot(X, weights)) I would guess the problem was new numpy outrunning theano compatibility.

I can make a PR in a day or two if no one has a better fix (@fonnesbeck / @twiecki I think you two present the code the most):

import pymc3 as pm
import numpy.random as npr
X = npr.randn(100, 1)
w = npr.randn(1)
y = X @ w + (npr.randn(100) * .1)

with pm.Model() as linear_model:
    weights = pm.Normal('weights', mu=0, sigma=1)
    noise = pm.Gamma('noise', alpha=2, beta=1)
    y_observed = pm.Normal('y_observed',
                mu=pm.math.dot(X, weights),
                sigma=noise,
                observed=y)

    posterior = pm.sample()

[fonnesbeck]

You can also just swap X and weights

weights.dot(X)

[ColCarroll]

Didn't realize this was 1-d. What about X * weights (also works), so that it doesn't make us linear algebra fiends sad?

[ahwillia]

Sorry I didn't mean to show the 1-d example -- the same error also happens for N-d. I was just curious if the error persisted for 1-d.

Swapping X and weights does seem to work. Why this is the case isn't clear to me...

Thanks for the quick responses!

[ColCarroll]

This was a good example! I suspect the problem is that something in theano breaks with new numpy arrays. weights.dot(X) returns a theano object, but X.dot(weights) returns a numpy object, I think.