repository·main·Indexed 25 days ago
https://github.com/sdatkinson/neural-amp-modelerA framework for training neural network models of audio amplifiers and exporting them to the specialized .nam file format. It includes tools for local training via the nam-full command, cloud training via Google Colab, and a detailed .nam file specification for storing model architecture, weights, and hardware calibration metadata.
Neural Amp Modeler (NAM) is a toolset used for training neural network models of amplifiers and exporting them into the .nam file format.
Note that this repository is specifically for the training and exporting workflow. If you want to play trained models in real time via a standalone application or plugin, you should use the NeuralAmpModelerPlugin repository instead.
Since .nam files are JSON-formatted, you can manually add or update calibration metadata using a text editor.
To update an existing model:
.nam file in a text editor."input_level_dbu""output_level_dbu"null.Warning: Always create a backup of your .nam file before editing.
Once the training process is complete (typically about 10 minutes):
.nam file..nam file to your local machine.The nam/models/wavenet module implements a slimmable class hierarchy that allows for dynamic adjustment of model size during inference or training. The top-level _WaveNet model manages an adjustment ratio (between 0 and 1) and propagates this scaling to its child modules using adjust_to() and adjust_to_random() methods.
Key scaling methods available via the _Slimmable mixin:
adjust_to(ratio): Sets the model to a specific size ratio.adjust_to_random(): Sets the model to a random size ratio.context_adjust_to_random(): Adjusts size within a specific context.To train a model, you need an input file (the standardized test signal) and an output file (the reamped signal from your gear).
If you want to skip reamping for your first model, you can use these pre-made files:
To find the analog voltage of your reamping send:
$$\text{dBu} = 20 \times \log_{10}\left(\frac{V_{\text{RMS}}}{0.7746}\right)$$
Example: 6.40 V RMS $\approx$ 18.3 dBu.
To build the project documentation locally, navigate to the docs directory and use the make command.
On Linux:
cd docs
make htmlOn Windows:
cd docs
make.bat htmlcd docs
make htmlPacked training allows training multiple compatible WaveNet submodels within a single larger masked WaveNet. This is useful for creating a single .nam file containing several WaveNet sizes that share the same temporal architecture but use different channel counts.
To use packed training, set the name in your model configuration to "PackedWaveNet" and run the nam-full command. No special CLI flags are required; the trainer automatically selects the PackedLightningModule based on the config.
Configuration Requirements:
submodels list contains individual WaveNet configurations.kernel_size, dilations, activation) must match across all submodels.channels) can differ between submodels.export.container_max_values to control the max_value thresholds in the exported SlimmableContainer. Options include "uniform" or a sorted list of values (one per submodel).{
"net": {
"name": "PackedWaveNet",
"config": {
"submodels": [
{
"name": "small",
"config": {
"layers_configs": [
{
"input_size": 1,
"condition_size": 1,
"channels": 3,
"head": {"out_channels": 1, "kernel_size": 1, "bias": true},
"kernel_size": 6,
"dilations": [1, 5, 29, 97, 227],
"activation": "LeakyReLU"
}
],
"head": null,
"head_scale": 0.01
}
},
{
"name": "large",
"config": {
"layers_configs": [
{
"input_size": 1,
"condition_size": 1,
"channels": 8,
"head": {"out_channels": 1, "kernel_size": 1, "bias": true},
"kernel_size": 6,
"dilations": [1, 5, 29, 97, 227],
"activation": "LeakyReLU"
}
],
"head": null,
"head_scale": 0.01
}
}
],
"export": {
"container_max_values": "uniform"
}
}
}
}To ensure accurate gain staging between digital models and analog source gear, you can define calibration levels in the NAM metadata. This involves measuring the relationship between the digital signal strength (dBFS) and the analog signal strength (dBu).
As of version 0.10.0, the NAM file specification supports two metadata fields:
input_level_dbu: The analog signal strength (in dBu) that corresponds to the loudest digital signal.output_level_dbu: The analog signal strength (in dBu) that corresponds to the loudest digital output.Providing these levels is optional. If omitted, models will still function, but gain staging may not accurately reflect the original analog hardware behavior.
Training requires two additional configuration files:
nam_full_configs/models/ (e.g., wavenet.json or wavenet_packed.json) to your desired filename (e.g., model.json).nam_full_configs/learning/ to your desired filename (e.g., learning.json).demo.json for a quick test run.default.json for standard use cases.default.json file.neural-amp-modeler package, use the Anaconda environment definitions located in the environments/ directory. For testing parity with GitHub Actions, refer to .github/workflows/python-pckage.yml to ensure your local environment matches the automated testing configuration.The nam-full command provides the full-featured option for training models locally. Ensure you have followed the installation instructions and activated your nam environment before running the command.
$ nam-full