Predictive and Deterministic tutorial

[mtsokol]

Hi @fritzo!

Sorry for a delay, here's a PR with a Predictive and deterministic tutorial discussed in #2803.

As proposed it's extended with a SVI method (with user defined guide and AutoGuide).

As you suggested, in the next PR I will add "Advanced Usage" section with poutine routines instead of Predictive class for further comparison.

Here are links to the notebook from the PR:

https://gist.github.com/mtsokol/fe00f78bb0b925f986d241622c2ec019
https://nbviewer.jupyter.org/gist/mtsokol/fe00f78bb0b925f986d241622c2ec019

What should I improve/modify?

My questions:

1. When performing inference with SVI it was failing with:

ValueError: Error while computing log_prob at site 'counts':
The value argument must be within the support

Although validate_args=False in GammaPoisson fixed it, actual values, which I was inspecting in a debugger, seemed within GammaPoissons support.

2. The guide I defined is rather arbitrary. Is it OK to use Delta this way or is it wrong?

Thank you for your help!

view jupyter notebook

[fritzo]

Hi @mtsokol this looks great, thanks for creating a tutorial! (and sorry for my delay in reviewing.) I have only minor comments:

• I didn't know about %config InlineBackend.figure_format='retina'. If this ends up looking good in your tutorial, we might want to change all other tutorials as well. (but let's check, since IIRC manually setting dpi=300 or svg output broke some tutorials in the past; we'll see on docs.pyro.ai once this merges).
• I think batch_size -> batch_shape and event_size -> event_shape would be more standard names.
• The model looks great, I like the use of sklearn's make_regression helper.
• Is there a reason you're using validate_args=False in the model? I'd rather encourage new users to use validation whenever possible. If you're simply trying to speed up MCMC, you might instead set jit_compile=True.
• nit: I think it would help to label both x and y axes on all figures. In particular, it took me a while to figure out what the "rate mean" and "rate std" plots show.
• Wherever you print(predictive.keys()) I think it would be helpful to additionally print sizes, say:

  for k, v in predictive.items():
      print(f"{k}: {tuple(v.shape)}")

[mtsokol]

@fritzo Thanks for a review!

Is there a reason you're using validate_args=False in the model?

Without validate_args=False SVI is failing with:

ValueError: Error while computing log_prob at site 'counts':
The value argument must be within the support

Here's a notebook with a full reproduction:
https://colab.research.google.com/drive/165osM9-jI2C8UbCppHTQe0ZpTc-S9WL-?usp=sharing

Is it a bug or my mistake in SVI part?

[fritzo]

Hmm, are your counts y_ actually integers? What if you

  y_ = torch.tensor((y**3)/100000. + 10., dtype=torch.float)
+ y_.round_().clamp_(min=0)

[mtsokol]

@fritzo Yes, it was my mistake. Now all working! I pushed changes with all fixes from the review.

Also gist with a current version:
https://gist.github.com/mtsokol/39f6d679834432bf29e1c35e79f547cf

[fritzo]

Looks great, thanks for creating this tutorial!