[chanshing]

Seeds all random number generators and enables deterministic PyTorch/cuDNN in the evaluation notebook so that repeated runs give the same results. Fixes the MAE table p-value, which tested the baseline's predictions against its own ground truth; it now runs a paired t-test on the per-participant errors of ActiNet against those of the baseline. The baseline random forest now uses sleep-transition handling in its HMM.

Breaking Changes

• build_pvalue_cell parameters are renamed from true_values/pred_values to values1/values2. Callers that pass them by keyword must be updated. Positional calls and the return format are unchanged.
• The "p-value" row returned by build_mae_table now has a different meaning: it compares the two models' errors instead of testing the baseline's predictions against the ground truth. Values will differ from earlier runs.

Files Changed

src/actinet/utils/eval_utils.py

• build_mae_table: the p-value is now a paired t-test between ActiNet's per-participant signed errors (true minus predicted) and the baseline's
• build_pvalue_cell: parameters are now generic, two paired samples instead of true and predicted values; the test itself is unchanged
• Formatting-only changes elsewhere

EvaluateModel.ipynb

• New seeding step that runs before evaluation. It seeds Python, NumPy and PyTorch (CPU and CUDA), sets PYTHONHASHSEED and the cuBLAS workspace config, enables torch.use_deterministic_algorithms, and disables cuDNN benchmarking
• Baseline random forest now runs with hmm_handle_sleep_transitions enabled
• The per-fold test participant list is now written to the versioned evaluation output directory, not a fixed outputs/actinet_vs_bbaa path
• Cells reformatted with Black; outputs remain cleared

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB