Function To Cast InferenceData Into tidy_draws Format
#36
Conversation
AFg6K7h4fhy2
added
feature
AFg6K7h4fhy2
added
Medium Priority
AFg6K7h4fhy2
modified the milestones:
[October 28, November 8],
[November 11, November 22]
AFg6K7h4fhy2
modified the milestones:
[November 11, November 22],
[November 25, December 6]
…nferencedata-into-tidy_draws-format
…nferencedata-into-tidy_draws-format
…nferencedata-into-tidy_draws-format
…nferencedata-into-tidy_draws-format
…nferencedata-into-tidy_draws-format
AFg6K7h4fhy2
modified the milestones:
[November 25, December 6],
[December 9, December 20]
…nferencedata-into-tidy_draws-format
…nferencedata-into-tidy_draws-format
|
Requested review from @dylanhmorris. Once a review is received, the author will request another review from Damon. Beyond the |
There was a problem hiding this comment.
Thanks @AFg6K7h4fhy2. Getting close but needs a few tweaks.
| ((pl.col(".chain") - 1) * pl.col("draws_per_chain")) | ||
| + pl.col(".iteration") | ||
| ).alias(".draw") | ||
| ) |
There was a problem hiding this comment.
Need to drop "draws_per_chain", but also it's not a given that all chains will have the same number of draws. Instead, more robust do compute this as .iteration + <n_draws_in_all_previous_chains>. Many ways to do that in polars.
|
|
||
| @pytest.fixture | ||
| def mock_inference_data(): | ||
| np.random.seed(42) |
There was a problem hiding this comment.
Running np.random.seed changes global state. Better practice to do something like this https://builtin.com/data-science/numpy-random-seed
| posterior_predictive = xr.Dataset( | ||
| { | ||
| "observed_hospital_admissions": ("chain", np.random.randn(2, 100)), | ||
| }, | ||
| coords={"chain": [0, 1]}, | ||
| ) |
There was a problem hiding this comment.
Why not run the test on the provided inference_data_1.nc? Or are you planning to remove it?
There was a problem hiding this comment.
I am on the fence about removing it. Seems good to have a canonical pyrenew-hew .nc file on hand esp. given that forecasttools-py does / will even more so interface abundantly with pyrenew models. On the other hand, having adequate and general idata / xarray representations seems good for testing too. I do not know if the latter must exist at the cost of the former. I lean towards having both, with the .nc file perhaps being used in notebooks and the "fake" idatas being used for testing.
| col in df.columns | ||
| for col in [".chain", ".draw", ".iteration", "variable", "value"] | ||
| ) | ||
|
|
There was a problem hiding this comment.
Would be good to check that individual values are as expected, not just that the draws are unique and one for each row.
…nferencedata-into-tidy_draws-format
Co-authored-by: Dylan H. Morris <[email protected]>
Co-authored-by: Dylan H. Morris <[email protected]>
…nferencedata-into-tidy_draws-format
| @@ -42,6 +43,8 @@ patsy = "^0.5.6" | |||
| nbformat = "^5.10.4" | |||
| nbclient = "^0.10.0" | |||
| jupyter = "^1.1.1" | |||
| pandas = "^2.2.3" | |||
| metaflow = "^2.13.9" | |||
…nferencedata-into-tidy_draws-format
For the scope of this PR, please refer to issue #18 .