NNSVS

https://nnsvs.github.io/
Admin
NNSVS is an open-source neural network-based singing voice synthesis library designed for research...
Tokens:176,312
Snippets:901
Trust Score:9.7
Update:1 month ago
Show doc for...
Tokens:
Raw
### example_xml_file function

Source: https://nnsvs.github.io/_modules/nnsvs/util.html

Retrieves the file path for an example XML file included in the package.

```python
def example_xml_file(key="haruga_kita"):
    """Get the path to an included xml file.

    Args:
        key (str): key of the file

    Returns:
        str: path to an example xml file

    Raises:
        FileNotFoundError: if the file is not found
    """
    return pkg_resources.resource_filename(__name__, join(EXAMPLE_DIR, f"{key}.xml"))
```

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

### example_xml_file(key='haruga_kita')

Source: https://nnsvs.github.io/modules/generated/nnsvs.util.example_xml_file.html

Returns the path to an included example XML file based on the provided key.

```APIDOC
## example_xml_file(key='haruga_kita')

### Description
Get the path to an included xml file.

### Method
Function

### Endpoint
`nnsvs.util.example_xml_file(key='haruga_kita')`

### Parameters
#### Path Parameters
- **key** (str) - Required - key of the file

### Returns
- **str** - path to an example xml file

### Raises
- **FileNotFoundError** - if the file is not found
```

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

### Install NNSVS for development

Source: https://nnsvs.github.io/_sources/installation.rst.txt

Clones the repository and installs the package in editable mode with development, linting, and testing dependencies.

```bash
git clone https://github.com/nnsvs/nnsvs.git && cd nnsvs
pip install -e ".[dev,lint,test]"
```

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

### Start MLflow UI

Source: https://nnsvs.github.io/_sources/optuna.rst.txt

Launch the MLflow UI to visualize and track hyperparameter search results.

```bash
mlflow ui --host 0.0.0.0
```

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

### nnsvs.util.example_xml_file

Source: https://nnsvs.github.io/modules/util.html

Returns the path to an example XML file included with the package.

```APIDOC
## nnsvs.util.example_xml_file

### Description
Get the path to an included xml file.

### Method
N/A

### Endpoint
N/A

### Parameters
None

### Request Example
N/A

### Response
#### Success Response
- **path** (string) - Path to the xml file.

#### Response Example
N/A
```

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

### Verify HnSincNSF import from command line

Source: https://nnsvs.github.io/_sources/train_vocoders.rst.txt

Verify the installation from the command line by importing and printing HnSincNSF. Should throw no errors if the installation is done properly.

```bash
python -c "from parallel_wavegan.models.nsf import HnSincNSF; print(HnSincNSF(1,1))"
```

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

### Install nnsvs if not already installed

Source: https://nnsvs.github.io/notebooks/Demos.html

Installs nnsvs from GitHub if the import fails. Uses %%capture to suppress output.

```python
%%capture
try:
    import nnsvs
except ImportError:
    ! pip install git+https://github.com/nnsvs/nnsvs

```

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

### Specify a custom model in a new file

Source: https://nnsvs.github.io/_sources/custom_models.rst.txt

Configuration example for a model defined in a custom file, such as nnsvs/test.py.

```yaml
netG:
_target_: nnsvs.test.MyModel
# the followings are arguments passed to your model's __init__ method
in_dim: 331
hidden_dim: 32
out_dim: 1
```

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

### Get note indices

Source: https://nnsvs.github.io/_modules/nnsvs/io/hts.html

Retrieves note start indices from HTS labels based on start times.

```python
def get_note_indices(labels):
    """Get note start indices from HTS labels

    Args:
        labels (nnmnkwii.io.hts.HTSLabelFile): HTS labels

    Returns:
        list: Note start indices
    """
    note_indices = [0]
    last_start_time = labels.start_times[0]
    for idx in range(1, len(labels)):
        if labels.start_times[idx] != last_start_time:
            note_indices.append(idx)
            last_start_time = labels.start_times[idx]
        else:
            pass
    return note_indices
```

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

### setup(config, device, collate_fn=collate_fn_default)

Source: https://nnsvs.github.io/modules/generated/nnsvs.train_util.setup.html

Sets up the training environment by creating the model, optimizer, learning rate scheduler, data loaders, tensorboard writer, logger, and scalers.

```APIDOC
## setup(config, device, collate_fn=collate_fn_default)

### Description
Sets up the training environment by creating the model, optimizer, learning rate scheduler, data loaders, tensorboard writer, logger, and scalers.

### Method
setup

### Parameters
- **config** (dict) - Required - Configuration for training.
- **device** (torch.device) - Required - Device to use for training.
- **collate_fn** (callable, optional) - Optional - Collate function. Defaults to collate_fn_default.

### Returns
- **tuple** - Tuple containing model, optimizer, learning rate scheduler, data loaders, tensorboard writer, logger, and scalers.
```

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

### get_note_indices(labels)

Source: https://nnsvs.github.io/modules/generated/nnsvs.io.hts.get_note_indices.html

Get note start indices from HTS labels. This function accepts an HTSLabelFile object and returns a list of note start indices.

```APIDOC
## get_note_indices(labels)

### Description
Get note start indices from HTS labels.

### Parameters
- **labels** (nnmnkwii.io.hts.HTSLabelFile) - Required - HTS labels

### Returns
- **list** - Note start indices
```

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

### Setup training environment

Source: https://nnsvs.github.io/_modules/nnsvs/train_util.html

Initializes logging, random seeds, CUDA settings, mixed precision, model, and optimizer for training.

```python
[docs]
def setup(config, device, collate_fn=collate_fn_default):
    """Setup for training

    Args:
        config (dict): configuration for training
        device (torch.device): device to use for training
        collate_fn (callable, optional): collate function. Defaults to collate_fn_default.

    Returns:
        (tuple): tuple containing model, optimizer, learning rate scheduler,
            data loaders, tensorboard writer, logger, and scalers.
    """
    if dist.is_initialized():
        rank = dist.get_rank()
        logger = getLogger(config.verbose) if rank == 0 else getLogger(0)
        sys.stdout = open(os.devnull, "w") if rank != 0 else sys.stdout
    else:
        logger = getLogger(config.verbose)
        rank = 0

    logger.info(OmegaConf.to_yaml(config))

    logger.info(f"PyTorch version: {torch.__version__}")

    if torch.cuda.is_available():
        from torch.backends import cudnn

        cudnn.benchmark = config.train.cudnn.benchmark
        cudnn.deterministic = config.train.cudnn.deterministic
        logger.info(f"cudnn.deterministic: {cudnn.deterministic}")
        logger.info(f"cudnn.benchmark: {cudnn.benchmark}")
        if torch.backends.cudnn.version() is not None:
            logger.info(f"cuDNN version: {torch.backends.cudnn.version()}")

    logger.info(f"Random seed: {config.seed}")
    init_seed(config.seed)

    if config.train.use_detect_anomaly:
        torch.autograd.set_detect_anomaly(True)
        logger.info("Set to use torch.autograd.detect_anomaly")

    if "use_amp" in config.train and config.train.use_amp:
        logger.info("Use mixed precision training")
        grad_scaler = GradScaler()
    else:
        grad_scaler = None

    # Model
    model = hydra.utils.instantiate(config.model.netG).to(device)
    logger.info(model)
    logger.info(
        "Number of trainable params: {:.3f} million".format(
            num_trainable_params(model) / 1000000.0
        )
    )

    # Distributed training
    if dist.is_initialized():
        device_id = rank % torch.cuda.device_count()
        model = DDP(model, device_ids=[device_id])

    # Optimizer
    optimizer_class = getattr(optim, config.train.optim.optimizer.name)
    optimizer = optimizer_class(
        model.parameters(), **config.train.optim.optimizer.params
    )

    # Scheduler
    lr_scheduler_class = getattr(
        optim.lr_scheduler, config.train.optim.lr_scheduler.name
    )
    lr_scheduler = lr_scheduler_class(
        optimizer, **config.train.optim.lr_scheduler.params
    )
```

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

### Install NNSVS on Google Colab

Source: https://nnsvs.github.io/_sources/installation.rst.txt

Checks for the existing installation and installs NNSVS from the GitHub repository if it is not found.

```python
%%capture
try:
   import nnsvs
except ImportError:
   ! pip install git+https://github.com/nnsvs/nnsvs
```

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

### Install NNSVS development dependencies

Source: https://nnsvs.github.io/_sources/optuna.rst.txt

Install the required development dependencies for NNSVS.

```bash
pip install -e ".[dev]"
```

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

### Run full recipe with CUDA_VISIBLE_DEVICES=0

Source: https://nnsvs.github.io/_sources/recipes.rst.txt

Runs the entire recipe from stage 0 to stage 99 on GPU 0. Omit CUDA_VISIBLE_DEVICES if only one GPU is available. Stage -1 is reserved for optional data download and may require manual download after signing a contract.

```bash
CUDA_VISIBLE_DEVICES=0 ./run.sh --stage 0 --stop-stage 99
```

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

### Install hyperparameter optimization requirements

Source: https://nnsvs.github.io/_sources/optuna.rst.txt

Explicitly install the packages required for hyperparameter optimization.

```bash
pip install mlflow optuna hydra-optuna-sweeper
```

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

### Pack models with custom models (stage 99)

Source: https://nnsvs.github.io/_sources/recipes.rst.txt

Run the model packaging step (stage 99) with custom model types. Specify the timelag, duration, and acoustic models as shown.

```bash
CUDA_VISIBLE_DEVICES=0 ./run.sh --stage 99 --stop-stage 99 \
        --timelag-model timelag_test \
        --duration-model duration_test \
        --acoustic-model acoustic_test
```

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

### Build documentation locally

Source: https://nnsvs.github.io/_sources/devdocs.rst.txt

Use sphinx-autobuild to generate and serve the documentation locally from the top of the nnsvs directory.

```bash
sphinx-autobuild docs docs/_build/html/
```

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

### Install NNSVS for inference

Source: https://nnsvs.github.io/_sources/installation.rst.txt

Installs the base NNSVS package for users who only need to run inference without training models.

```bash
pip install nnsvs
```

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

### Verify HnSincNSF import in Jupyter

Source: https://nnsvs.github.io/_sources/train_vocoders.rst.txt

Verify the installation by importing HnSincNSF in a Jupyter notebook. Should throw no errors if the installation is done properly.

```python
In [1]: from parallel_wavegan.models.nsf import HnSincNSF
In [2]: HnSincNSF(1,1)
```

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

### Example of conf directory structure

Source: https://nnsvs.github.io/recipes.html

Shows the directory layout for model-specific configurations in a recipe, including subdirectories for duration, timelag, and acoustic models, each with data, model, and train configuration files.

```text
conf
├── train
│   ├── duration
│   │   ├── data
│   │   │   └── myconfig.yaml
│   │   ├── model
│   │   │   └── duration_mdn.yaml
│   │   └── train
│   │       └── myconfig.yaml
│   └── timelag
│       ├── data
│       │   └── myconfig.yaml
│       ├── model
│       │   └── timelag_mdn.yaml
│       └── train
│           └── myconfig.yaml
└── train_acoustic
    └── acoustic
        ├── data
        │   └── myconfig.yaml
        ├── model
        │   └── acoustic_resf0convlstm.yaml
        └── train
            └── myconfig.yaml

```

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

### Fix HTS label offset to zero

Source: https://nnsvs.github.io/_modules/nnsvs/io/hts.html

The fix_label_offset_to_zero function adjusts the start and end times of HTS labels so that the first label starts at zero.

```python
def fix_label_offset_to_zero(labels):
    """Fix label offset to zero

    Args:
        labels (nnmnkwii.io.hts.HTSLabelFile): HTS labels

    Returns:
        nnmnkwii.io.hts.HTSLabelFile: HTS labels with fixed offset
    """
    offset = labels.start_times[0]
    if offset > 0:
        labels.start_times = np.asarray(labels.start_times) - offset
        labels.end_times = np.asarray(labels.end_times) - offset
    return labels
```

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

### Pack models (stage 99)

Source: https://nnsvs.github.io/_sources/recipes.rst.txt

Run the model packaging step (stage 99) without custom models. This packs the trained models into a single directory.

```bash
CUDA_VISIBLE_DEVICES=0 ./run.sh --stage 99 --stop-stage 99
```

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

### Setup GAN Training Environment

Source: https://nnsvs.github.io/_modules/nnsvs/train_util.html

Configures the training environment for GAN models, including logging, device settings, mixed precision, and model instantiation for both generator and discriminator.

```python
def setup_gan(config, device, collate_fn=collate_fn_default):
    """Setup for training GAN

    Args:
        config (dict): configuration for training
        device (torch.device): device to use for training
        collate_fn (callable, optional): collate function. Defaults to collate_fn_default.

    Returns:
        (tuple): tuple containing model, optimizer, learning rate scheduler,
            data loaders, tensorboard writer, logger, and scalers.
    """
    if dist.is_initialized():
        rank = dist.get_rank()
        logger = getLogger(config.verbose) if rank == 0 else getLogger(0)
        sys.stdout = open(os.devnull, "w") if rank != 0 else sys.stdout
    else:
        logger = getLogger(config.verbose)
        rank = 0

    logger.info(OmegaConf.to_yaml(config))

    logger.info(f"PyTorch version: {torch.__version__}")

    if torch.cuda.is_available():
        from torch.backends import cudnn

        cudnn.benchmark = config.train.cudnn.benchmark
        cudnn.deterministic = config.train.cudnn.deterministic
        logger.info(f"cudnn.deterministic: {cudnn.deterministic}")
        logger.info(f"cudnn.benchmark: {cudnn.benchmark}")
        if torch.backends.cudnn.version() is not None:
            logger.info(f"cuDNN version: {torch.backends.cudnn.version()}")

    logger.info(f"Random seed: {config.seed}")
    init_seed(config.seed)

    if config.train.use_detect_anomaly:
        torch.autograd.set_detect_anomaly(True)
        logger.info("Set to use torch.autograd.detect_anomaly")

    if "use_amp" in config.train and config.train.use_amp:
        logger.info("Use mixed precision training")
        grad_scaler = GradScaler()
    else:
        grad_scaler = None

    # Model G
    netG = hydra.utils.instantiate(config.model.netG).to(device)
    logger.info(netG)
    logger.info(
        "[Generator] Number of trainable params: {:.3f} million".format(
            num_trainable_params(netG) / 1000000.0
        )
    )

    if dist.is_initialized():
        device_id = rank % torch.cuda.device_count()
        netG = DDP(netG, device_ids=[device_id])

    # Optimizer and LR scheduler for G
    optG, schedulerG = _instantiate_optim(config.train.optim.netG, netG)

    # Model D
    netD = hydra.utils.instantiate(config.model.netD).to(device)
    logger.info(netD)
    logger.info(
        "[Discriminator] Number of trainable params: {:.3f} million".format(
            num_trainable_params(netD) / 1000000.0
        )
    )

    if dist.is_initialized():
        device_id = rank % torch.cuda.device_count()
        netD = DDP(netD, device_ids=[device_id])

    # Optimizer and LR scheduler for D
    optD, schedulerD = _instantiate_optim(config.train.optim.netD, netD)

    # DataLoader
    data_loaders, samplers = get_data_loaders(config.data, collate_fn, logger)

    set_epochs_based_on_max_steps_(
        config.train, len(data_loaders["train_no_dev"]), logger
    )

    # Resume
    _resume(logger, config.train.resume.netG, netG, optG, schedulerG)
    _resume(logger, config.train.resume.netD, netD, optD, schedulerD)

    if config.data_parallel:
        netG = nn.DataParallel(netG)
        netD = nn.DataParallel(netD)
```

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

### Install parallel_wavegan nnsvs branch

Source: https://nnsvs.github.io/_sources/train_vocoders.rst.txt

Install the nnsvs branch of the parallel_wavegan repository to enable neural vocoder support in NNSVS. This branch includes code for NSF.

```bash
pip install git+https://github.com/r9y9/ParallelWaveGAN@nnsvs
```

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

### Install NNSVS in editable mode

Source: https://nnsvs.github.io/_sources/devdocs.rst.txt

Install the full requirements with editable mode enabled to make local changes available to the Python environment without manual re-installation.

```bash
pip install -e ".[dev,lint,test,docs]"
```

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

### Run stage 7 to prepare post-filter features

Source: https://nnsvs.github.io/_sources/train_postfilters.rst.txt

Run stage 7 of the recipe to prepare input and output features for training post-filters. Requires a trained acoustic model; the acoustic model's checkpoint directory is specified with --acoustic-model.

```bash
CUDA_VISIBLE_DEVICES=0 ./run.sh --stage 7 --stop-stage 7 \
    --acoustic-model acoustic_test
```

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

### Initialize Training Components

Source: https://nnsvs.github.io/_modules/nnsvs/train_util.html

Initializes data loaders, model, optimizer, and logging infrastructure. Handles configuration for MLflow or Tensorboard and loads input/output scalers.

```python
    # DataLoader
    data_loaders, samplers = get_data_loaders(config.data, collate_fn, logger)

    set_epochs_based_on_max_steps_(
        config.train, len(data_loaders["train_no_dev"]), logger
    )

    # Resume
    _resume(logger, config.train.resume, model, optimizer, lr_scheduler)

    if config.data_parallel:
        model = nn.DataParallel(model)

    # Mlflow
    if config.mlflow.enabled:
        mlflow.set_tracking_uri("file://" + get_original_cwd() + "/mlruns")
        mlflow.set_experiment(config.mlflow.experiment)
        # NOTE: disable tensorboard if mlflow is enabled
        writer = None
        logger.info("Using mlflow instead of tensorboard")
    else:
        # Tensorboard
        if rank == 0:
            writer = SummaryWriter(to_absolute_path(config.train.log_dir))
        else:
            writer = None

    # Scalers
    if "in_scaler_path" in config.data and config.data.in_scaler_path is not None:
        in_scaler = joblib.load(to_absolute_path(config.data.in_scaler_path))
        in_scaler = MinMaxScaler(
            in_scaler.min_, in_scaler.scale_, in_scaler.data_min_, in_scaler.data_max_
        )
    else:
        in_scaler = None
    if "out_scaler_path" in config.data and config.data.out_scaler_path is not None:
        out_scaler = joblib.load(to_absolute_path(config.data.out_scaler_path))
        out_scaler = StandardScaler(
            out_scaler.mean_, out_scaler.var_, out_scaler.scale_
        )
    else:
        out_scaler = None

    return (
        model,
        optimizer,
        lr_scheduler,
        grad_scaler,
        data_loaders,
        samplers,
        writer,
        logger,
        in_scaler,
        out_scaler,
    )
```

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

### PredictionType.PROBABILISTIC usage

Source: https://nnsvs.github.io/_modules/nnsvs/base.html

Example usage for MDN-based models.

```python
# training
mdn_params = model(x)
# inference
mu, sigma = model.inference(x)
```

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

### PredictionType.DIFFUSION usage

Source: https://nnsvs.github.io/_modules/nnsvs/base.html

Example usage for diffusion model predictions.

```python
# training
noise, x_recon = model(x)

# inference
y = model.inference(x)
```

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

### List parallel_wavegan configurations

Source: https://nnsvs.github.io/_sources/train_vocoders.rst.txt

View the available model configuration files in the parallel_wavegan directory.

```bash
$ tree conf/parallel_wavegan
conf/parallel_wavegan
├── hn-sinc-nsf_sr24k_pwgD.yaml
├── hn-sinc-nsf_sr48k_hifiganD.yaml
├── hn-sinc-nsf_sr48k_pwgD.yaml
└── hn-sinc-nsf_sr48k_pwgD_test.yaml
```

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

### PredictionType.DETERMINISTIC usage

Source: https://nnsvs.github.io/_modules/nnsvs/base.html

Example usage for non-MDN single-stream models.

```python
# training
y = model(x)
# inference
y = model.inference(x)
```
Tokens:
Raw