Symbolic shape argument to Distribution

[bwengals]

For example, the model

with pm.Model() as model:
    a = pm.Normal("a", mu=0.0, sd=1.0, shape=tt.as_tensor_variable(3))

will fail with exception:

----------------------------------------------------------
TypeError                Traceback (most recent call last)
<ipython-input-6-40fdcd9234e8> in <module>()
      1 with pm.Model() as model:
----> 2     a = pm.Normal("a", mu=0.0, sd=1.0, shape=tt.as_tensor_variable(3))

~/pymc3/pymc3/distributions/distribution.py in __new__(cls, name, *args, **kwargs)
     33                 raise TypeError("observed needs to be data but got: {}".format(type(data)))
     34             total_size = kwargs.pop('total_size', None)
---> 35             dist = cls.dist(*args, **kwargs)
     36             return model.Var(name, dist, data, total_size)
     37         else:

~/pymc3/pymc3/distributions/distribution.py in dist(cls, *args, **kwargs)
     44     def dist(cls, *args, **kwargs):
     45         dist = object.__new__(cls)
---> 46         dist.__init__(*args, **kwargs)
     47         return dist
     48 

~/pymc3/pymc3/distributions/continuous.py in __init__(self, mu, sd, tau, **kwargs)
    253         assert_negative_support(tau, 'tau', 'Normal')
    254 
--> 255         super(Normal, self).__init__(**kwargs)
    256 
    257     def random(self, point=None, size=None, repeat=None):

~/pymc3/pymc3/distributions/distribution.py in __init__(self, shape, dtype, defaults, *args, **kwargs)
    145             dtype = theano.config.floatX
    146         super(Continuous, self).__init__(
--> 147             shape, dtype, defaults=defaults, *args, **kwargs)
    148 
    149 

~/pymc3/pymc3/distributions/distribution.py in __init__(self, shape, dtype, testval, defaults, transform, broadcastable)
      51         self.shape = np.atleast_1d(shape)
     52         if False in (np.floor(self.shape) == self.shape):
---> 53             raise TypeError("Expected int elements in shape")
     54         self.dtype = dtype
     55         self.type = TensorType(self.dtype, self.shape, broadcastable)

TypeError: Expected int elements in shape

This situation arises for me when working with GP classes. Would it be a good idea to handle a symbolic shape argument given to a distribution?

[twiecki]

Would be a great idea and we have discussed it before but I think it might be quite difficult.

[bwengals]

Ah ok, ill look into it

[ferrine]

The current idea is not possible. Because we need fixed latent dimensionality for for inference approaches MCMC and VI. Flexibility can be introduced when doing AEVB only.

[aseyboldt]

I thought about more flexible shape handling a bit, as it came up a couple of times now. I've got a few ideas, and a sketch of an implementation for some of it. I'll push it later, maybe we can mention it in the meeting tomorrow.