### Start NAM Full-Featured Trainer

Source: https://github.com/sdatkinson/neural-amp-modeler/blob/main/docs/source/tutorials/full.md

Execute this command in the terminal to launch the full-featured NAM trainer after installation.

```console
$ nam-full
```

--------------------------------

### Install neural-amp-modeler

Source: https://github.com/sdatkinson/neural-amp-modeler/blob/main/docs/source/installation.md

Install the neural-amp-modeler package using pip after setting up PyTorch.

```bash
$ pip install neural-amp-modeler
```

--------------------------------

### Install GPU PyTorch

Source: https://github.com/sdatkinson/neural-amp-modeler/blob/main/docs/source/installation.md

Install a GPU-compatible version of PyTorch. This command is for CUDA 12.9. Refer to the PyTorch website for the most up-to-date instructions.

```bash
$ pip install torch --index-url https://download.pytorch.org/whl/cu129
```

--------------------------------

### Update neural-amp-modeler

Source: https://github.com/sdatkinson/neural-amp-modeler/blob/main/docs/source/installation.md

Update an existing installation of the neural-amp-modeler package to the latest version.

```bash
pip install --upgrade neural-amp-modeler
```

--------------------------------

### Slimmable Training Class Hierarchy

Source: https://github.com/sdatkinson/neural-amp-modeler/blob/main/nam/models/wavenet/slimmable.md

This diagram illustrates the inheritance structure of slimmable modules, starting from nn.Module and extending to specialized layers like _SlimmableConv1dBase and its subclasses.

```text
nn.Module
├── _Slimmable (ABC, nn.Module)
│   │   Mixin: manages _adjust_size ratio (0–1); adjust_to(), adjust_to_random(), context_adjust_to_random()
│   │
│   ├── _WaveNet (_Slimmable, nn.Module)
│   │   Top-level model; propagates adjust_to() to child modules
│   │
│   └── _SlimmableConv1dBase (nn.Conv1d, _Slimmable)
│       Base: abstract _get_adjusted_weight_and_bias(); forward() uses sliced w,b when adjusted
│       │
│       ├── _SlimmableRechannelIn
│       │   Rechannel into layer array. First layer: slice out only; later: slice in and out.
│       │   _max_adjust_size = out_channels
│       │
│       ├── _SlimmableConvLayer
│       │   Layer conv: channels → mid_channels. Gated: mid=2*ch, slice w[:2*adj,:adj,:]
│       │   _max_adjust_size = in_channels
│       │
│       ├── _SlimmableInputMixer
│       │   Condition → mid_channels. Slice output only.
│       │   _max_adjust_size = out_channels
│       │
│       ├── _Slimmable1x1
│       │   1×1 in residual path. Slice both in and out (in_channels == out_channels).
│       │   _max_adjust_size = in_channels (= out_channels)
│       │
│       └── _SlimmableHeadRechannel
│           channels → 1. Slice input only.
│           _max_adjust_size = in_channels
```

--------------------------------

### Check PyTorch GPU Availability

Source: https://github.com/sdatkinson/neural-amp-modeler/blob/main/docs/source/installation.md

Verify if PyTorch can detect and use the NVIDIA GPU. If this returns 'False', your PyTorch installation needs to be fixed.

```python
import torch; print(torch.cuda.is_available())
```

--------------------------------

### Check PyTorch GPU Version

Source: https://github.com/sdatkinson/neural-amp-modeler/blob/main/docs/source/installation.md

Confirm that the installed PyTorch version supports the GPU. A version string containing 'cu' indicates GPU support.

```python
import torch; print(torch.__version__)
```

--------------------------------

### Build Documentation Locally (Linux)

Source: https://github.com/sdatkinson/neural-amp-modeler/blob/main/README.md

Use this command to build the documentation locally on a Linux system. Ensure you are in the 'docs' directory.

```bash
cd docs
make html
```

--------------------------------

### Run NAM Full-Featured Training

Source: https://github.com/sdatkinson/neural-amp-modeler/blob/main/docs/source/tutorials/full.md

Initiate the training process by providing the paths to your data, model, and learning configuration files, along with the desired output directory.

```console
nam-full \
path/to/data.json \
path/to/model.json \
path/to/learning.json \
path/to/outputs
```

--------------------------------

### Build Documentation Locally (Windows)

Source: https://github.com/sdatkinson/neural-amp-modeler/blob/main/README.md

Use this command to build the documentation locally on a Windows system. Ensure you are in the 'docs' directory.

```bat
cd docs
make.bat html
```

--------------------------------

### Run Packed Training with Full Trainer

Source: https://github.com/sdatkinson/neural-amp-modeler/blob/main/docs/source/tutorials/packed-training.md

Use the normal full trainer command with 'PackedWaveNet' in the model config. No extra command-line flags are needed.

```console
$ nam-full path/to/data.json path/to/model.json path/to/learning.json path/to/outputs
```

--------------------------------

### Launch NAM GUI Trainer

Source: https://github.com/sdatkinson/neural-amp-modeler/blob/main/docs/source/tutorials/gui.md

Run this command in your terminal to launch the graphical user interface for training NAM models.

```console
$ nam
```

--------------------------------

### Configure Data Path for Training

Source: https://github.com/sdatkinson/neural-amp-modeler/blob/main/docs/source/tutorials/full.md

Edit the data configuration file to specify the paths to your input and output audio files. Ensure the 'delay' parameter accurately reflects the latency in samples between the two files.

```json
"common": {
    "x_path": "C:\\path\\to\\input.wav",
    "y_path": "C:\\path\\to\\output.wav",
    "delay": 0
}
```

--------------------------------

### Uninstall PyTorch Components

Source: https://github.com/sdatkinson/neural-amp-modeler/blob/main/docs/source/installation.md

Uninstall PyTorch and related libraries. This is a step before reinstalling to fix GPU issues.

```bash
$ pip uninstall torch torchvision torchaudio
```

--------------------------------

### Packed WaveNet Model Configuration

Source: https://github.com/sdatkinson/neural-amp-modeler/blob/main/docs/source/tutorials/packed-training.md

Configure packed training by specifying 'PackedWaveNet' as the model name and defining submodels with their respective configurations. Temporal settings must match across submodels, while channel counts can differ.

```json
{
  "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"
      }
    }
  }
}
```

--------------------------------

### Helper Function for Channel Calculation

Source: https://github.com/sdatkinson/neural-amp-modeler/blob/main/nam/models/wavenet/slimmable.md

This helper function calculates the number of channels based on a given ratio and maximum size, ensuring a minimum of 1 channel.

```python
_ratio_to_channels(ratio, max_size) → 1 + min(floor(ratio * max_size), max_size - 1)
```

--------------------------------

### Slice Semantics for Weight Tensors

Source: https://github.com/sdatkinson/neural-amp-modeler/blob/main/nam/models/wavenet/slimmable.md

This table details how weight tensors (w) and biases (b) are sliced for different slimmable classes, specifying the slicing logic and any relevant notes for each.

```text
Class                   | Slicing                               | Notes                         
|-------------------------|---------------------------------------|-------------------------------
| _SlimmableRechannelIn   | first: w[:adj,:,:]; later: w[:adj,:adj,:] | Rechannel into layer array   
|_SlimmableConvLayer     | gated: w[:2*adj,:adj,:], b[:2*adj]; else: w[:adj,:adj,:] | Dilated conv, optional gate 
|_SlimmableInputMixer    | w[:adj,:,:], b[:adj]                  | Condition mixer              
|_Slimmable1x1           | w[:adj,:adj,:], b[:adj]               | Residual path 1×1            
|_SlimmableHeadRechannel | w[:,:adj,:], b unchanged              | Head: channels → 1           
```

--------------------------------

### Data Flow in Slimmable Mode

Source: https://github.com/sdatkinson/neural-amp-modeler/blob/main/nam/models/wavenet/slimmable.md

This diagram outlines the data flow through the slimmable components of the WaveNet model, showing how data is processed by rechanneling layers, main layers, and the head rechanneling layer.

```text
_LayerArray (slimmable=True)
│
├── _rechannel:     _SlimmableRechannelIn (input_size → channels)
│
├── _layers[i]:     _Layer
│   ├── _conv:      _SlimmableConvLayer (channels → mid_channels; gated)
│   ├── _input_mixer: _SlimmableInputMixer (condition_size → mid_channels)
│   └── _layer1x1:  _Slimmable1x1 (bottleneck → channels)
│
└── _head_rechannel: _SlimmableHeadRechannel (channels → ``head.out_channels``; kernel_size 1 only for slimmable)
```

=== COMPLETE CONTENT === This response contains all available snippets from this library. No additional content exists. Do not make further requests.