Shape error in random method for HalfNormal

[junpenglao]

with pm.Model() as m:
    a = pm.HalfNormal('a', np.ones((3, 1)), shape=(3, 10))

a.random(point=m.test_point)

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-35-a1da77da6c66> in <module>()
----> 1 a.random(point=m.test_point)

~/Documents/pymc3/pymc3/model.py in __call__(self, *args, **kwargs)
     40 
     41     def __call__(self, *args, **kwargs):
---> 42         return getattr(self.obj, self.method_name)(*args, **kwargs)
     43 
     44 

~/Documents/pymc3/pymc3/distributions/continuous.py in random(self, point, size)
    749         return generate_samples(stats.halfnorm.rvs, loc=0., scale=sd,
    750                                 dist_shape=self.shape,
--> 751                                 size=size)
    752 
    753     def logp(self, value):

~/Documents/pymc3/pymc3/distributions/distribution.py in generate_samples(generator, *args, **kwargs)
    522             dist_shape: {dist_shape}
    523             broadcast_shape: {broadcast_shape}
--> 524         '''.format(size=size, dist_shape=dist_shape, broadcast_shape=broadcast_shape))
    525 
    526     # reshape samples here

TypeError: Attempted to generate values with incompatible shapes:
            size: 1
            dist_shape: (3, 10)
            broadcast_shape: (3, 1)

I have a vague feeling this is something to do with broadcasting sigma parameter in Normal/HalfNormal, I think I have seen it before...

[junpenglao]

Same error for Normal:

with pm.Model() as m:
    a = pm.Normal('a', 0, np.ones((3, 1)), shape=(3, 10))
a.random(m.test_point)

also

with pm.Model() as m:
    a = pm.Normal('a', np.zeros((3, 1)), 1, shape=(3, 10))
a.random(m.test_point)

So it is a general broadcast error

[junpenglao]

Maybe in random method we can broadcast the numpy value that feed into generate_samples? @ColCarroll

[ColCarroll]

I am pretty sure if shape is supplied, I tried to make sure it was equal to np.broadcast(*inputs).shape. In this case, the inputs were np.zeros((3, 1)) and 1 (the default standard deviation).

It sounds like you're suggesting that we set a.shape = np.broadcast(shape, *inputs).shape. I think that would work, but it is a little funny, since the result of that broadcast might be different from what the user intended.

[junpenglao]

I meant something like:

return generate_samples(stats.halfnorm.rvs, loc=0., scale=np.broadcast(self.shape, sd),
                          dist_shape=self.shape,
                          size=size)

[springcoil]

Has there been work on this?

[twiecki]

I don't think so, although there were many shape fixes so worth checking if the behavior is still the same.

[lucianopaz]

Closed via #3456