Juniper NAM Trainer

A free trainer for Neural Amp Modeler that runs in Google Colab. It makes standard NAM A2 .nam files that load in any up-to-date NAM player, trained the way I train the captures in Juniper Studio's plugins.

It's built on the official NAM trainer, with my training changes added on top: early stopping, an ESR auto-stop, and an extra high-frequency term in the loss.

Open the trainer in Google Colab

Free. You need a Google account. Colab's free GPU works.

What it does

  • Standard NAM A2 models. By default each .nam file holds a full model and a lite model, so players that support it show the slim (size) option. You can also train full only or lite only.
  • Batch training. Train a whole folder of captures, A to Z or Z to A. Each model is saved the moment it finishes, so a disconnect never loses one, and captures that are already done get skipped.
  • Stops at the right time. Training stops when the model stops improving, or as soon as it reaches your ESR target (0.0001 by default, or turn it off). The best version is the one that gets saved, not the last one.
  • Honest ESR. The ESR is measured on NAM's held-out section, audio the model never trains on, and the full and lite models are scored separately.
  • Latency is found automatically. It uses NAM's own calibration on the standard input file, or cross-correlation on your own DI.
  • NAM's recording checks on the standard input file, with an option to skip them.
  • Any input file. Use the standard NAM input (built in) or your own DI. Files that aren't 48 kHz are resampled, and stereo is folded to mono.
  • Your files come from Google Drive or are uploaded from your computer, and models are saved back to Drive or downloaded.

What's saved in every .nam file

  • Calibration levels: input and output dBu at 0 dBFS, read from the file name (_in_12.0_out_6.5) or typed in. You can turn calibration off.
  • Loudness: measured the standard NAM way, so players can level-match it with other models.
  • Gear info: name, modeled by, gear make, model and type, and tone type.
  • A training report: the held-out ESR of the full and lite models, the best epoch, how many epochs ran and why training stopped, the latency used, the recording-check result, and the trainer and NAM versions.
  • The sample rate (48 kHz) and the date it was trained.

How to use it

  1. Reamp your gear with the standard NAM input file (or your own DI) at 48 kHz, and export each take as a .wav the same length as the input.
  2. To train from Google Drive, put the captures in a folder in your Drive (or in a captures subfolder of it) and type that folder's name in the trainer's settings. Models are saved to an output subfolder. Or upload the files straight from your computer.
  3. Open the trainer, then go to Runtime › Change runtime type and pick a GPU.
  4. Run the cells in order. Settings are a simple form. When the last cell finishes, your .nam files are in the output folder.

Name captures like My Amp Gain 5_in_12.0_out_6.5.wav to save the calibration levels with the model: _in_ is the input dBu at 0 dBFS and _out_ is the output dBu at 0 dBFS.

Questions or problems? Get in touch.

Built on Neural Amp Modeler by Steven Atkinson (MIT licence).