AutoGuide does not support sequential pyro.plate

Misc.
#1

I want to use SVI to inference for a multivariate sequence with shape [365,5]. The code can be simplified as:

sigma=pyro.sample(dist.uniform(0,1).expand([5]).to_event(1))
#X[365,5],Y[365,5]
for i in pyro.plate(“data”,X):
…obs=pyro.sample(‘obs_{}’.format(i),dist.Normal(loc=X[:,i],scale=sigma[i]),obs=Y[:,i])

X and Y have a one-year length with 5 features. I gave a different sigma of Normal distribution for each feature. But this setting can not run. I appreciate any help. The error is :

ValueError: at site “obs_0”, invalid log_prob shape
Expected [], actual [365]
Try one of the following fixes:

  • enclose the batched tensor in a with plate(…): context
  • .to_event(…) the distribution being sampled
  • .permute() data dimensions
#2

I also try the below codes but still has problems. AutoDiagonalNormal is used in the SVI. AutoGuide does not support sequential pyro.plate. Same problems also can be seen https://forum.pyro.ai/t/times-series-model-getting-notimplementederror-autoguidelist-does-not-support-sequential-pyro-plate/854 but haven’t clearly solved. The code is :

for i in pyro.plate('time',time_len):
    for j in pyro.plate('feature_%d'%(i),feature_len):
        y_obs=Y[i,j]
        obs=pyro.sample('obs_%d_%d' % (i, j), Normal(loc=X[i, j], scale=sigma[j]),
                        obs=y_obs)

The error is:

NotImplementedError: AutoGuide does not support sequential pyro.plate
Trace Shapes:
 Param Sites:
Sample Sites:
#3

You could try replacing the sequential pyro.plate() with a simple for loop. This should work well as long as all your sample sites are reparameterized.

for i in range(time_len):
    for j in range(feature_len):
        y_obs=Y[i,j]
        obs=pyro.sample('obs_%d_%d' % (i, j), Normal(loc=X[i, j], scale=sigma[j]),
                        obs=y_obs)