Repository navigation
Releases
ParaNAM Trainer 0.7.6
ParaNAM Trainer 0.7.6
-
Plan → Rig controls adds Plan rules. Select controls and use Avoid all to skip settings where every selected control meets the condition.
-
Knobs have editable Avoid < and Avoid > thresholds. Older projects carry their endpoint exclusions forward as equivalent thresholds.
-
Plan edits reuse compatible recorded takes. Other takes stay in the project folder and can be reused if they match a later plan.
-
Capture highlights the controls that differ from the last recorded take, including after selecting another row or returning to the page. Confirm My rig matches… to stop the highlighting.
-
Scrolling over numeric value fields no longer changes their values accidentally.
-
Panel dividers are aligned across the Trainer interface.
-
A dimmed overlay with an animated spinner appears as soon as a project starts loading. It also blocks input during longer calculations and stays visible briefly between consecutive operations to avoid flicker.
-
Projects can be renamed or duplicated from the Projects page. Duplicates receive the first available numbered suffix.
Full Changelog: v0.7.5...v0.7.6
ParaNAM Trainer 0.7.5
What's new
Knob values in Plan, Capture and Compare
- Positive values on bipolar knobs now display a leading
+when the control's range crosses zero (minimum below zero and maximum above zero). - The rule applies to the knob readouts in Plan, Capture and Compare, to the small minimum and maximum labels beside the knobs in Plan and Capture, and to the Control settings column of the Capture table.
- Zero, negative values and strictly unipolar controls retain their previous notation. Only the displayed text changes; numeric parameter values, model files and audio processing are unaffected.
Compare
- Model player is now up to date with the last changes made in the plugin code, these include optimization to the knobs smoothing process and fix of the bug relative to the Output mode.
Main window
- Project header: the open project's name is displayed larger and in white, while the separator dot and the Automatically saved status keep their previous size and muted color.
- Full-width grey separators now extend from the border of the left navigation panel to the right edge of the window. The project name, pages and Back/Next controls retain their 30px side padding, preserving the resizable Capture table.
Plan, Capture, Queue and Compare
- Plan control cards follow the pointer during drag-and-drop, while the other cards animate into position. Hover highlights control cards and both the border and icon of the add card.
- Selecting a control skips recalculating capture recommendations. Capture grid limits account for both marked Gain/Drive controls.
- Queue fits its columns on first opening and then preserves their widths. Queue and Capture reuse existing table cells, and Compare keeps the model selector intact when its contents are unchanged.
- Scrolling over closed dropdowns no longer changes their selection.
Update notification
- GitHub-style Markdown release notes are rendered directly in the update window, including headings, bold and italic text, lists, links, code blocks and tables.
- Resizable notes area: when the update dialog grows or shrinks, the release notes panel changes height while the version notification at the top stays fixed.
- Visible Windows installer: an update started from the notification opens the normal installer interface rather than silently hiding the installation steps. Existing verification and safe shutdown checks are preserved.
Automatic cleanup of old installers
- On startup, the Trainer removes all completed or abandoned updater download folders, including installers left behind by earlier releases, not just the most recent update.
- This clears downloaded setup executables, DMGs, Debian packages, portable archives and their associated temporary files, logs and manifests.
- Safety: an updater directory whose installation result is still pending is left untouched. Files locked by the operating system are skipped and tried again on a later launch. Training projects, recordings, exports and unrelated files are never targeted.
Full Changelog: v0.7.4...v0.7.5
ParaNAM Trainer 0.7.4
ParaNAM Trainer 0.7.4
- Plan: Rig controls now use the same control cards as Capture. Cards wrap onto new rows, can be reordered by dragging, and have a red × for removal. Use the grey dashed + card to add a new control.
- Control settings: Edit settings inline with automatic saving. Switch positions have editable cards. Dragging a position changes the option order used for training and keeps the default on the same named position.
- Plan and Capture: Knobs show their minimum and maximum. The red arc starts at zero when zero is in range, or at the minimum otherwise, and extends to the current position.
- Capture and Train: Capture keeps the audio stream open between takes. Train metadata stays editable while recording and edits are saved when the capture finishes.
- Queue: Column widths follow their contents. Validation Full ESR is now Val. ESR.
- Window: ParaNAM remembers the previous window geometry, and main panels expand horizontally with the window.
- Export: Parametric switches use
optionswith a stable neuralvalueand a separate displayindex_sw, including the Full/Lite branches. Static NAM containers and their submodels use format version0.7.0.
Full Changelog: 0.7.3...v0.7.4
ParaNAM Trainer 0.7.3
ParaNAM Trainer 0.7.3
- Static capture: choose Record or Load WAV for a named capture. Load the single aligned mono PCM return of the input selected in Plan (e.g. the TONE3000 sweep or TTS v11), with the same sample rate and length. The last nine seconds of that sweep are held out for validation, so no separate validation WAV is needed. Imported returns retain their original level and are labelled as not timing-verified.
- Settings: instrument-input calibration is global and stays in place when opening projects. Historical capture and job calibration snapshots remain intact. ASIO devices have a Settings button above Refresh to open the selected driver's control panel.
- Train: remembers the last author for projects that do not already specify one. Input impedance accepts Ω, kΩ and MΩ with either decimal separator. Export writes the optional numeric input_z in ohms.
- Plan and Capture: Avoid min also skips nonzero minimum positions. Capture diagrams show larger values and only the selected switch label. Every cell in the running capture row is bright green.
- Recording integrity: reported stream loss aborts capture immediately. A stopped or stalled stream also fails the capture instead of accepting a partial return.
- PNAM compatibility: ParametricSlimmableContainer still includes Full and Lite. This release writes legacy switch enum_names in neural order, including exports from resumed jobs. Existing indexed files remain readable; index_sw output is deferred.
- Compare: uses the plugin's current single-network DSP: weight updates at a maximum of 10 per second, with a 100 ms linear, sample-by-sample ramp in the convolution kernels while controls move.
All five native packages were rebuilt, tested, and checked from commit e3c1ee7102cc.
Full Changelog: v0.7.2...0.7.3
ParaNAM Trainer 0.7.2
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
HyperA2models. - Existing
ParametricOutputGainwrappers.
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
olderHyperA2andParametricOutputGainfiles. - 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
ParaNAM Trainer 0.7.1
-
Changed text color in Captures and Queue lists to be more vibrant
-
Fixed deactivate "Pre-warm audio interface" not working and renamed this function to a more appropriate "Check timing stability before each capture"
Full Changelog: v0.7.0...v0.7.1
ParaNAM Trainer 0.7.0
ParaNAM Trainer 0.7.0 preview
- Complex Switch keeps the existing plan for each joint switch state. Simple
Switch adds one discrete Latin-hypercube dimension, contributing to suggested
capture counts like one knob. Only actual switch positions are captured. Both
modes use the same categorical neural-model encoding. - Knobs can concentrate captures in the lower or upper half of their range.
A mirrored exponential sampling transform places about 73% of initial draws
in the selected half before capture-grid rounding and duplicate removal.
Physical knob values and model normalization remain unchanged. - Avoid max excludes the upper endpoint from LHS and corner captures, mirroring
Avoid zero. Clean output volume requires its maximum as the reference and
therefore cannot use Avoid max or Gain/Drive. - Clean output volume trains the neural effect at maximum volume and measures
a separate attenuation curve with short validation-signal recordings at the
selected knob step (default 1.000 when Clean volume is enabled). All other
controls stay at maximum. Peak and RMS attenuation are compared against the
volume-10 reference; a difference greater than 0.2 dB shows an advisory warning
with Dismiss and Go back to Plan actions. The output curve is fitted
from the measured RMS levels.
Maximum-volume references are recorded first. A silent minimum is measured as
a noise-floor clip, without requiring timing alignment or a loopback. Audible
captures retain timing checks; an explicitly configured loopback must work.
At least one other neural control is required. - Renaming a control updates project parameter references while retaining its
audio files and plan. Controls can be reordered in the Plan page. Capture
displays the Clean volume split as “Volume”. - Compare model and impulse response lists extend to the bottom of the page and
allow selected entries to be removed. Bundled impulse responses are protected. - Clean volume is stored in a ParametricOutputGain wrapper and applied after
the neural model using a monotone PCHIP curve and 5 ms smoothing. Compare,
NAM Parametric Plugin 0.4.5 preview and NeuralAudio 0.1.4 preview support this
format. Older players reject it; volume-only changes do not regenerate neural
weights. Static export in the new plugin includes the selected volume gain. - Adds Truncated (Phillip Self): the original 38-second inputTrunc.wav plus
the bundled NAM standard first 10 and last 9 seconds (57 seconds total).
The original 38 seconds remain unchanged and form the training segment;
the validation segments stay excluded from fitting. Separate held-out captures
continue to use the existing short validation signal. Source:
https://github.com/phillipmself/neural-amp-modeler-parametric/releases/tag/training-inputs-v1 - LR cadence selectors now sit beside their respective decay fields, with
aligned numeric fields and transparent row containers. - Capture progress resets when cancelled. Capture and Queue rows use state
colors, their lists have a shaded “#” corner header, and Capture labels its
first column Type. A system notification sound plays when a capture is
saved successfully. Queue no longer shows a redundant order column. - Capture can pre-warm the interface with 250 ms of low-level white noise and
monitor the Return signal through ASIO Monitor outputs when enabled. - Scrollbar tracks are hidden. Training log horizontal position is preserved
while new text arrives, and remaining-time estimates use every observed batch
and stay visible while paused. - The project name is bold in the header, Train shows metadata before settings,
its neural-network icon uses filled nodes, and Compare removes the two status
lines shown beneath the player. - Parametric exports include Train ESR Full/Lite and the saved advanced settings
in training metadata. Legacyvalidation_esr,validation_reference, and
validation_esr_meanfields are removed. - Held-out validation recommendation is min(max(8, 4 × dimensions), 24) per
Complex Switch state; Simple Switch dimensions count alongside knobs. The
initial validation count is 12. ESR_worst reports the worst complete held-out
setting alongside mean ESR for Full/Lite. - Removes the draft 0.6.3 beta Cosine schedules. Existing ordinary-cadence jobs
retain resume compatibility; beta Cosine jobs require a new training.
Existing projects and recordings are not silently regenerated when new planning features are selected.
Full Changelog: v0.1.0...v0.7.0
ParaNAM Trainer 0.6.3 beta
ParaNAM Trainer 0.6.3 beta
This beta is prepared on beta/0.6.3-cosine-validation. The release remains a
draft; 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.
Draft installers: Windows x64, Linux x64 DEB/portable archive, macOS Apple Silicon/Intel.
ParaNAM Trainer 0.6.2 Preview
ParaNAM Trainer 0.6.2
- Capture advances to the next recording with its progress bar reset to 0%.
Only controls whose positions changed pulse slowly toward light grey. Selecting
a capture manually or confirming that the rig matches stops the pulse. - Capture switches now use the same compact slide-switch drawing as Compare,
retaining Capture's tile backgrounds and colors. Up to two knobs can be marked
Gain/Drive; corner planning crosses both gain ranges with the other controls'
corners and switch combinations. The corners checkbox has a clearer name. - Compare opens without a project. Its model controls keep a fixed minimum gap
and wrap into as many columns as fit. Models and IRs are shown in two adjacent
lists at the bottom. The Instrument meter has the width of one Monitor channel. - Numeric fields and knob values accept either a decimal point or comma.
Typed model knob values continue to apply on Enter or loss of focus. - Replaced both LR preset menus with independent cadence selectors beside each
learning rate: per epoch and per 62 optimizer updates (A2 cadence).
Both decay multipliers remain editable. Legacy editor values are represented
by an equivalent average decay; newly queued/Retrain jobs use that representation.
Existing jobs remain resumable: resume preserves their frozen original
scheduler configuration, optimizer state and checkpoint counters. Equivalent
average decay does not imply identical discrete reduction instants for a new job.
Normal recording/architecture/evaluation compatibility checks still apply. - New parametric projects and the HyperA2 preset use Adam, batch 16, LR 0.004,
hypernetwork LR 0.001, decay 0.994/0.998497 per 62 updates, MRSTFT enabled,
NY 8192, MRSTFT weight 0.0005, weight decay 0.000000317, rank-2 low-rank
hypernetwork deltas, shared dilated convolution kernels and summed Full/Lite
losses. Clipping, NaN/Inf gradient checks, cuDNN benchmark and deterministic
training default off. HyperWaveNet is unchanged except cuDNN benchmark is off.
Existing projects retain their settings; hypernetwork fields stay hidden in static mode. - Default epoch budgets are 50 for parametric and 200 for static training.
Switching capture type remembers each budget separately. - Queue now allows every row to move up or down, including running and finished
jobs. The running process remains associated with its instance. Removed the
paused-job batch/device editors; above-list actions align right except Start/Pause
queue. Renamed the actions to Pause selected training and Cancel selected training. - Elapsed time advances with a monotonic clock, independently of worker updates.
Elapsed and remaining time retain seconds after one hour. - Added Open project folder for the selected project on the Projects page.
Installers: Windows x64, Linux x64 DEB/portable archive, macOS Apple Silicon/Intel.
ParaNAM Trainer 0.6.1 Preview
ParaNAM Trainer 0.6.1
- Restored the original update/control cadences and use this fixed order in both
LR selectors: A2 Cadence, A2 Slow, A2 Slower, A2 Dataset-normalized √,
A2 Dataset-normalized, then Manual. Follow is no longer a selectable mode.
Intervals are respectively62,62 × F / sqrt(2),62 × F,
62 × sqrt(S/62), andSoptimizer updates, where
F = max(1, knobs) × 2^switchesand S is complete updates per training epoch.
A2 Slow keeps its previous formula; it does not use sqrt(F). Initial LR values
remain editable and new-project defaults remain A2 Dataset-normalized. - LR Decay and Hypernetwork LR Decay are editable with every preset. Selecting a
cadence keeps the entered decay; preview and training use each group's value
rather than forcing 0.994. The default remains 0.994. Values 0 and 1 are supported.
Legacy Follow configurations display their explicit effective cadence and decay;
unchanged legacy checkpoint identities are preserved, while any separate edit
creates independently scheduled settings. - Removed the 0.6.0 block on jobs using the earlier preset IDs. IDs, effective
settings, frozen scheduler intervals, optimizer state and checkpoint counters
are preserved on resume. No training is automatically restarted or migrated.
Normal capture, architecture and evaluation-protocol checks remain in force. - Fixed the paused-queue/Retrain path: starting a queue can again resume the older
paused instance instead of being blocked before a newly retrained job can run.
Retrain restores the effective instance settings rather than stale editor values. - Full and Lite ESR charts now start with the same fixed logarithmic range:
0.005 at the bottom, 0.5 at the top. The range remains unchanged while scores fit
inside it; only an exceeded side expands by logarithmic decades. Empty charts,
zero ESR and nonfinite scores remain safe. The X axis still spans all planned epochs. - Includes the state recovery, None audio routes, static-mode field visibility,
and verified GitHub update notifications introduced in previous releases.
Installers: Windows x64, Linux x64 DEB/portable archive, macOS Apple Silicon/Intel.