Stabilize synthetic stats fixtures used by training tests#1232
Merged
Conversation
The synthetic stats files used by the coupled/ace training integration tests were populated with per-variable `np.random.randn()` draws for both means and stds. Stds drawn near zero (or negative) caused normalized inputs to blow up, which the CRPS + NoiseConditionedSFNO training case amplified through multi-step rollouts to produce outputs in the 1e+3 to 1e+17 range and occasional NaN values during inference. Iteration over `set(...)` of variable names made the per-variable std assignment depend on `PYTHONHASHSEED`, so the same test would land on a pathological std on some CI invocations but not others. Changes: - `fme.ace.testing.fv3gfs_data.get_scalar_dataset` / `save_scalar_netcdf`: replace `randn()` default with `center + uniform(-0.05, 0.05)` so per-variable noise stays bounded. - `fme.ace.testing.save_stats_netcdfs`: new helper that writes a paired mean/std file with the std centered at 1.0, so callers cannot forget to keep stds bounded away from zero. - Route `StatsData`, `fme.coupled.data_loading.test_data_loader. create_coupled_data_on_disk`, and `fme.ace.test_train._setup` through the new helper. Drop the now-redundant `stats_std_fill_value` parameter from `_setup`. Resolves #1143
AnnaKwa
approved these changes
Jun 8, 2026
AnnaKwa
left a comment
AnnaKwa
left a comment
Contributor
There was a problem hiding this comment.
Thanks for tracking this down!
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes flakiness in `fme/coupled/test_train.py::test_train_and_inference[3-True]` (and the GPU-skipped `test_train_and_inference_with_derived_forcings`) by removing a hidden source of pathological inputs in the synthetic stats fixtures.
The integration tests built per-variable normalization stats by drawing `np.random.randn()` for both means and stds. Stds occasionally landed near zero or negative, which made normalized inputs blow up. The CRPS + `NoiseConditionedSFNO` training case amplified that through multi-step rollouts and produced inference outputs in the `1e+3` to `1e+17` range and occasional NaN values. Variation between runs came from `list(set(...))` of variable names depending on `PYTHONHASHSEED`, so the same test would pull a pathological std assignment on some CI invocations but not others. Verified locally on CPU: five runs of the CRPS case with the old behavior produced `PRESsfc` magnitudes ranging from `1e+0` to `3e+3` (one run with std drawn at `-1.2`); after the fix all five runs stayed at `~1e+0` with no NaN/Inf.
Changes:
`fme.ace.testing.fv3gfs_data.get_scalar_dataset` / `save_scalar_netcdf`: replace the `randn()` default with `center + uniform(-0.05, 0.05)` so per-variable noise stays bounded.
`fme.ace.testing.save_stats_netcdfs`: new helper that writes a paired mean/std file with the std centered at 1.0, so callers cannot forget to keep stds bounded away from zero.
Route `StatsData`, `fme.coupled.data_loading.test_data_loader.create_coupled_data_on_disk`, and `fme.ace.test_train._setup` through the new helper. Drop the now-redundant `stats_std_fill_value` parameter from `_setup`.
Tests added (existing integration tests continue to pass; the fix is on the test fixture itself)
If dependencies changed, "deps only" image rebuilt and "latest_deps_only_image.txt" file updated
Resolves #1143