Variational Autoencoders Example is broken

[gioelelm]

Issue Description

I tried to run the Variational Autoencoders tutorial here and it returns an error

Environment

Both on my machine (Macbook Pro 2019 Catalina) and a Google Colab environment.

pyro.__version__=='1.8.1'
torch.__version__=='1.11.0+cu113'

I tried a couple of combinations of torch and pyro versions, I got always the dame bug

Error

ValueError: Error while computing log_prob at site 'obs':
Expected value argument (Tensor of shape (256, 784)) to be within the support (Boolean()) of the distribution Bernoulli(probs: torch.Size([256, 784])), but found invalid values:
tensor([[0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.],
        ...,
        [0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.]])
        Trace Shapes:            
         Param Sites:            
 decoder$$$fc1.weight 400  50    
   decoder$$$fc1.bias     400    
decoder$$$fc21.weight 784 400    
  decoder$$$fc21.bias     784    
        Sample Sites:            
          latent dist 256   |  50
                value 256   |  50
             log_prob 256   |    
             obs dist 256   | 784
                value 256   | 784

[eb8680]

Hmm, I can't reproduce this on Colab, and I would expect this warning/error to be suppressed by the disabling of distribution runtime checks pyro.distributions.enable_validation(False). Can you run pip freeze and post the output?

[GStechschulte]

I ran this example locally (2019 MacBook Pro using macOS Monterey) a couple days ago and it worked fine:

pyro-api=0.1.2
pyro-ppl=1.8.1
torch=1.11.0

My results were in line with the example notebook results.