sampling with Poisson emissions gives ValueError

[bantin]

The following lines give a value error most of the time:

true_lds = ssm.LDS(15, 10, dynamics="none", emissions="poisson")
x,y = true_lds.sample(1000)

The traceback looks like this:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-13-27cbe882d496> in <module>
----> 1 x,y = true_lds.sample(1000)

/Users/bantin/Documents/Linderman-Shenoy/ssm/ssm/lds.py in sample(self, T, input, tag, prefix, with_noise)
   1018
   1019     def sample(self, T, input=None, tag=None, prefix=None, with_noise=True):
-> 1020         (_, x, y) = super().sample(T, input=input, tag=tag, prefix=prefix, with_noise=with_noise)
   1021         return (x, y)

/Users/bantin/Documents/Linderman-Shenoy/ssm/ssm/lds.py in sample(self, T, input, tag, prefix, with_noise)
    252         # Sample observations given latent states
    253         # TODO: sample in the loop above?
--> 254         y = self.emissions.sample(z, x, input=input, tag=tag)
    255         return z[pad:], x[pad:], y[pad:]
    256

/Users/bantin/Documents/Linderman-Shenoy/ssm/ssm/emissions.py in sample(self, z, x, input, tag)
    640         z = np.zeros_like(z, dtype=int) if self.single_subspace else z
    641         lambdas = self.mean(self.forward(x, input, tag))
--> 642         y = npr.poisson(lambdas[np.arange(T), z, :])
    643         return y
    644

~/anaconda3/lib/python3.7/site-packages/autograd/tracer.py in f_wrapped(*args, **kwargs)
     46             return new_box(ans, trace, node)
     47         else:
---> 48             return f_raw(*args, **kwargs)
     49     f_wrapped.fun = f_raw
     50     f_wrapped._is_autograd_primitive = True

mtrand.pyx in numpy.random.mtrand.RandomState.poisson()

common.pyx in numpy.random.common.disc()

common.pyx in numpy.random.common.discrete_broadcast_d()

common.pyx in numpy.random.common.check_array_constraint()

ValueError: lam value too large

I think this means that somewhere the lambda parameter passed to npr.Poisson is too large. According to the documentation, "Because the output is limited to the range of the C long type, a ValueError is raised when lam is within 10 sigma of the maximum representable value."

I don't think this is a big issue per se -- it's just because the forward function generating the lambdas is unstable. But maybe the default initializations could be adjusted so that this doesn't happen.

[bantin]

Possible fix in 0d6aa48. Will close once merged.

[bantin]

• Also should switch default link to softplus

[bantin]

Switched default link to softplus in f363c8a