Unable to reproduce result of AIR example

Hi there,

I am running the AIR example code with the exact same parameters as provided here but unable to reproduce the result of the progression of ELBO as shown in the tutorial (Fig 3, Left). The reconstruction seems fine and the count accuracy is pretty much the same, but in my experiments the ELBO has consistently converge to around 620.0 instead of above 650.0 as shown in the figure. This is kinda odd to me. Is it possible that the figure is actually showing the likelihood instead of ELBO? Or perhaps some other reason could cause the discrepancy?


from the tutorial:

We used Pyro commit c0b38ad with PyTorch 0.2.0.post4 . Inference ran for approximately 4 hours on an NVIDIA K80 GPU. (Note that even though we set the random seed, this isn’t sufficient to make inference deterministic when using CUDA.)

there are numerous possible reasons for this - if you’d like to reproduce it exactly, those versions are your best bet. though if your count accuracy and reconstruction are the same, then the elbo isn’t really important.