Repository navigation
ParaNAM Trainer 0.6.3 beta
Pre-release
Pre-release
What's new
This beta is prepared on beta/0.6.3-cosine-validation. The public 0.6.2 release and main branch are unchanged.
- Added Cosine (whole budget) independently for the audio network and
hypernetwork. It warms up linearly from zero to the configured initial/peak LR
over the first 2% of optimizer updates, then follows a cosine curve to 1% of
that LR at the end of the epoch budget. The budget uses actual complete
training batches after validation exclusion, context and per-capture drop-last.
Every rate is computed directly from the base LR and saved step counter.
Both budgets and counters are frozen in the job/checkpoint and restored on
pause/resume, without re-estimating the schedule. Existing epoch/A2 cadences
remain available. The decay field is inactive for a group using cosine. - Increased recommended held-out validation captures to
min(max(8, 4 × knobs), 24)per joint switch combination. The default
n_validationfor new projects is 12. Saved plans keep their previous counts;
captures remain distinct from training settings. Switch-only models retain
their existing behavior because there are no unrecorded control settings to
hold out after all switch states have been captured. Coarse knob grids must
have enough distinct settings for both train and validation counts. - Added ESR_worst, the highest ESR of a complete held-out control setting.
It pools audio energies within each setting and takes the maximum across
settings, so quiet failures cannot be hidden by louder recordings. Logged
alongside mean ESR for Full, Lite and the combined validation score, and saved
in progress/history/checkpoints. It reuses validation predictions without an
extra inference pass. Missing capture identities are not reported as a false
worst-setting score; nonfinite setting scores remain visible as failures. - Best-checkpoint selection continues to use mean Full validation ESR. Existing
jobs retain their original scheduler and validation plan on resume. Use a new
training instance to try cosine; changing the number of held-out settings
requires updating the capture plan and recording those captures.
Full Changelog: v0.6.2-preview...v0.6.3-beta