Skip to content

ParaNAM Trainer 0.7.2

Choose a tag to compare

@DLC86 DLC86 released this 09 Nov 17:00
· 362 commits to main since this release

Capture planning

Four Latin Hypercube Sampling methods

The Plan page now offers four reproducible ways to distribute training
captures across the control space:

  • Standard — the original seeded LHS, unchanged and still the default.
  • Maximin — chooses the candidate LHS design with the greatest minimum
    separation between its points.
  • Orthogonal — strength-2 orthogonal-array LHS, for sample counts and
    dimensions satisfying its mathematical constraints.
  • Near-orthogonal — minimizes correlations between sampled dimensions
    without the stricter Orthogonal sample-count requirements.

The plan's seed and selected method are saved with the project. The default
anchor, sampling bias, control-step quantization, categorical switches and
corner captures can reduce the ideal properties of an initial LHS design.
Existing plans and recorded captures are not silently regenerated.

Alternative validation selection: farthest from train

The Validation selection setting in Plan offers two methods:

  • Independent LHS (current) — the existing held-out, independently seeded
    LHS split, still selected by default.
  • Farthest from train (maximin) — first builds the training LHS, then
    greedily picks validation combinations from a reproducible, quantized
    candidate pool. Each new point maximizes its distance from the nearest
    training/default/corner point or previously selected validation point.
    Controls are normalized by their range; Simple Switch states are treated
    as categories and balanced, while Complex Switch combinations retain
    separate coverage.

This optional validation method probes less densely covered regions of the
control space. It is not an unbiased estimate of average accuracy over
all possible settings. Changing the selection requires regenerating the plan
and capturing the newly planned settings. Old projects retain Independent LHS.

Updated capture-count recommendations

Recommended training LHS density rises from 15 to 16 captures per knob or
Simple Switch dimension
, with separate coverage for Complex Switch
combinations. The validation-count recommendation is unchanged. Existing
saved capture counts remain as they were.

Default parametric input and epoch recommendations

New Parametric projects now select Truncated (Phillip Self) as
their initial training input, with 200 suggested epochs. Selecting
Standard or TTS input v11 suggests 50 epochs instead. The
suggestion updates when switching input types only while the previous
suggested budget is still in use: manually edited epoch counts are
preserved. Static projects continue to default to Standard input and
200 epochs, and the last input choice for each capture type is
remembered separately. Existing capture plans and saved training
instances are not silently changed.

ParametricSlimmableContainer model format

New dual-size parametric exports use the top-level architecture
ParametricSlimmableContainer instead of HyperA2, retaining
the same Lite/Full HyperWaveNet networks, neural control definitions,
weight ordering and file schema version 1.0.0
.

Measured Clean/Linear Volume attenuation is exported as an optional
extensions.linear_volume PCHIP curve at the root of the model,
outside config.params. The volume control therefore does not
become an input to the hypernetwork; the runtime applies its smoothed
attenuation after the neural signal.

Updated NAM Parametric Plugin and NeuralAudio can load:

  • The new container, with or without extensions.linear_volume.
  • Existing HyperA2 models.
  • Existing ParametricOutputGain wrappers.

The plugin continues to support static baking with both model sizes
and the chosen volume level. Compatibility is backward-reading,
not forward-reading:
older players that dispatch only HyperA2
will reject the newly named container, even when no Linear Volume
extension is present. Update the player before opening new exports.

Training

Alternative checkpoint selection for HyperA2

Advanced Training Settings → Checkpoint selection now provides three
independent Full/Lite checkpoint selection rules:

  • Lowest Validation ESR (current) — the original behavior, kept as default.
  • Validation within 5%, then lowest Train ESR — among checkpoints within
    5% of that size's final lowest Validation ESR, select the lowest Train ESR
    measured on held-out audio from training-control combinations. Checkpoint
    candidates that cannot win are pruned.
  • Combined ESR — choose the checkpoint minimizing
    alpha × Validation ESR + (1 − alpha) × Train ESR.
    The configurable Checkpoint alpha ranges from 0 to 1 and defaults
    to 0.75. It is enabled only for Combined ESR.

Full and Lite can be selected from different epochs, including in exports
and resumable snapshots. Neither optional criterion changes gradient
updates, the optimizer or learning-rate schedule. Existing parametric jobs
keep their original selection setting; static training is unchanged.

Improvements and fixes

Selected text colors in Captures and Queue

Selecting a row now preserves the original text colors throughout the
selected row, rather than replacing them with white. Running text stays
green, completed/captured text cyan and warning text yellow. The row's
selection background is explicitly rendered in neutral grey to avoid
system-specific blue highlights on Windows, Linux or macOS.

Clean Volume: silence means no detected test signal

A muted minimum-volume recording is now recognized by the absence of
the played timing-test signal
in the amplifier return, rather than
an arbitrary RMS/peak noise-floor threshold. Background hiss does not
by itself mean the test signal was detected.

  • A verified silent minimum no longer produces a misleading
    peak-versus-RMS non-linearity warning.
  • Verified muted minima map to exactly zero output gain, not the
    measured background-noise RMS.
  • Earlier captures with saved raw timing recordings can be rechecked
    and obsolete warnings cleared without re-recording.
  • Non-silent recordings retain the existing 0.20 dB discrepancy check.
  • Loopback timing verification remains intact; missing expected timing
    signals at other knob settings still fail.

Windows auto-update preserves installer choices

Removed the empty /TASKS= argument from the Windows auto-updater.
Inno Setup can now retain the user's previous installer task selections
instead of forcibly clearing them during a silent update. This includes
the desktop shortcut and optional NVIDIA GPU support. NVIDIA remains an
opt-in choice for new manual installations; existing GPU libraries are
not removed by an update that does not reinstall them.

Compatibility

  • Existing projects, captures and exported models remain supported.
  • Independent validation LHS and the original lowest-Validation-ESR
    checkpoint rule remain the defaults.
  • The HyperA2 Full/Lite neural topology and weights and static A2
    formats are unchanged. New parametric file exports use
    ParametricSlimmableContainer; updated players keep reading
    older HyperA2 and ParametricOutputGain files.
  • Selecting new validation points requires a new capture plan; changing
    checkpoint selection requires a new training instance with matching
    settings for later resume.

Full Changelog: v0.7.1...v0.7.2