Skip to content

Conversation

@ordabayevy
Copy link
Member

  • This updates partial_unroll to support unrolling of Markov dimensions (only history=1).
  • Markov variable names have to follow the var_suffix naming convention (e.g., ("x_0", "x_prev", "x_curr")).
  • _expected_modified_partial_sum_product is removed.

@ordabayevy
Copy link
Member Author

#402 (comment) same issue here.

@fehiepsi
Copy link
Member

@ordabayevy The issue is fixed now in the master branch.

Copy link
Member

@eb8680 eb8680 left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This looks good to me, nice simplification! I just have one nit. (Sorry for the delayed review, I was pretty busy with NeurIPS last week.)

Currently only plates with history={0, 1} are supported.
Markov vars are assumed to have names that follow ``var_suffix`` formatting
(e.g., ``("x_0", "x_prev", "x_curr")``).
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can you add an assertion that checks this property of plate_to_step?

@ordabayevy
Copy link
Member Author

@eb8680 no worries at all! The assertion checks that var prefixes match between var_prev and var_curr names (e.g., x in x_prev and x_curr). I'll add analogous assertion for init names (x_0) once init will be included in the step.

@eb8680 eb8680 merged commit 6fa4e3f into pyro-ppl:master Dec 14, 2020
@ordabayevy
Copy link
Member Author

Thanks for reviewing!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants