NeMo Models#


NeMo models contain everything needed to train and reproduce Conversational AI models:

  • neural network architectures

  • datasets/data loaders

  • data preprocessing/postprocessing

  • data augmentors

  • optimizers and schedulers

  • tokenizers

  • language models

NeMo uses Hydra for configuring both NeMo models and the PyTorch Lightning Trainer.


Every NeMo model has an example configuration file and training script that can be found here.

The end result of using NeMo, Pytorch Lightning, and Hydra is that NeMo models all have the same look and feel and are also fully compatible with the PyTorch ecosystem.


NeMo comes with many pretrained models for each of our collections: ASR, NLP, and TTS. Every pretrained NeMo model can be downloaded and used with the from_pretrained() method.

As an example, we can instantiate QuartzNet with the following:

import nemo.collections.asr as nemo_asr

model = nemo_asr.models.EncDecCTCModel.from_pretrained(model_name="QuartzNet15x5Base-En")

To see all available pretrained models for a specific NeMo model, use the list_available_models() method.


For detailed information on the available pretrained models, refer to the collections documentation:


NeMo leverages PyTorch Lightning for model training. PyTorch Lightning lets NeMo decouple the conversational AI code from the PyTorch training code. This means that NeMo users can focus on their domain (ASR, NLP, TTS) and build complex AI applications without having to rewrite boiler plate code for PyTorch training.

When using PyTorch Lightning, NeMo users can automatically train with:

  • multi-GPU/multi-node

  • mixed precision

  • model checkpointing

  • logging

  • early stopping

  • and more

The two main aspects of the Lightning API are the LightningModule and the Trainer.

PyTorch Lightning LightningModule#

Every NeMo model is a LightningModule which is an nn.module. This means that NeMo models are compatible with the PyTorch ecosystem and can be plugged into existing PyTorch workflows.

Creating a NeMo model is similar to any other PyTorch workflow. We start by initializing our model architecture, then define the forward pass:

class TextClassificationModel(NLPModel, Exportable):
    def __init__(self, cfg: DictConfig, trainer: Trainer = None):
        """Initializes the BERTTextClassifier model."""
        super().__init__(cfg=cfg, trainer=trainer)

        # instantiate a BERT based encoder
        self.bert_model = get_lm_model(

        # instantiate the FFN for classification
        self.classifier = SequenceClassifier(
def forward(self, input_ids, token_type_ids, attention_mask):
    No special modification required for Lightning, define it as you normally would
    in the `nn.Module` in vanilla PyTorch.
    hidden_states = self.bert_model(
        input_ids=input_ids, token_type_ids=token_type_ids, attention_mask=attention_mask
    logits = self.classifier(hidden_states=hidden_states)
    return logits

The LightningModule organizes PyTorch code so that across all NeMo models we have a similar look and feel. For example, the training logic can be found in training_step:

def training_step(self, batch, batch_idx):
    Lightning calls this inside the training loop with the data from the training dataloader
    passed in as `batch`.
    # forward pass
    input_ids, input_type_ids, input_mask, labels = batch
    logits = self.forward(input_ids=input_ids, token_type_ids=input_type_ids, attention_mask=input_mask)

    train_loss = self.loss(logits=logits, labels=labels)

    lr = self._optimizer.param_groups[0]['lr']

    self.log('train_loss', train_loss)
    self.log('lr', lr, prog_bar=True)

    return {
        'loss': train_loss,
        'lr': lr,

While validation logic can be found in validation_step:

def validation_step(self, batch, batch_idx):
    Lightning calls this inside the validation loop with the data from the validation dataloader
    passed in as `batch`.
    if self.testing:
        prefix = 'test'
        prefix = 'val'

    input_ids, input_type_ids, input_mask, labels = batch
    logits = self.forward(input_ids=input_ids, token_type_ids=input_type_ids, attention_mask=input_mask)

    val_loss = self.loss(logits=logits, labels=labels)

    preds = torch.argmax(logits, axis=-1)

    tp, fn, fp, _ = self.classification_report(preds, labels)

    return {'val_loss': val_loss, 'tp': tp, 'fn': fn, 'fp': fp}

PyTorch Lightning then handles all of the boiler plate code needed for training. Virtually any aspect of training can be customized via PyTorch Lightning hooks, Plugins, callbacks, or by overriding methods.

For more domain-specific information, see:

  • Automatic Speech Recognition (ASR)

  • Natural Language Processing (NLP)

  • Text-to-Speech Synthesis (TTS)

PyTorch Lightning Trainer#

Since every NeMo model is a LightningModule, we can automatically take advantage of the PyTorch Lightning Trainer. Every NeMo example training script uses the Trainer object to fit the model.

First, instantiate the model and trainer, then call .fit:

# We first instantiate the trainer based on the model configuration.
# See the model configuration documentation for details.
trainer = pl.Trainer(**cfg.trainer)

# Then pass the model configuration and trainer object into the NeMo model
model = TextClassificationModel(cfg.model, trainer=trainer)

# Now we can train with by calling .fit

# Or we can run the test loop on test data by calling

All trainer flags can be set from from the NeMo configuration.


Hydra is an open-source Python framework that simplifies configuration for complex applications that must bring together many different software libraries. Conversational AI model training is a great example of such an application. To train a conversational AI model, we must be able to configure:

  • neural network architectures

  • training and optimization algorithms

  • data pre/post processing

  • data augmentation

  • experiment logging/visualization

  • model checkpointing

For an introduction to using Hydra, refer to the Hydra Tutorials.

With Hydra, we can configure everything needed for NeMo with three interfaces:

  • Command Line (CLI)

  • Configuration Files (YAML)

  • Dataclasses (Python)


NeMo provides YAML configuration files for all of our example training scripts. YAML files make it easy to experiment with different model and training configurations.

Every NeMo example YAML has the same underlying configuration structure:

  • trainer

  • exp_manager

  • model

Model configuration always contain train_ds, validation_ds, test_ds, and optim. Model architectures vary across domains, therefore, refer to the ASR, NLP, and TTS Collections documentation for more detailed information on Model architecture configuration.

A NeMo configuration file should look similar to the following:

# PyTorch Lightning Trainer configuration
# any argument of the Trainer object can be set here
    devices: 1 # number of gpus per node
    accelerator: gpu
    num_nodes: 1 # number of nodes
    max_epochs: 10 # how many training epochs to run
    val_check_interval: 1.0 # run validation after every epoch

# Experiment logging configuration
    exp_dir: /path/to/my/nemo/experiments
    name: name_of_my_experiment
    create_tensorboard_logger: True
    create_wandb_logger: True

# Model configuration
# model network architecture, train/val/test datasets, data augmentation, and optimization
        manifest_filepath: /path/to/my/train/manifest.json
        batch_size: 256
        shuffle: True
        manifest_filepath: /path/to/my/validation/manifest.json
        batch_size: 32
        shuffle: False
        manifest_filepath: /path/to/my/test/manifest.json
        batch_size: 32
        shuffle: False
        name: novograd
        lr: .01
        betas: [0.8, 0.5]
        weight_decay: 0.001
    # network architecture can vary greatly depending on the domain

More specific details about configuration files for each collection can be found on the following pages:

NeMo ASR Configuration Files


With NeMo and Hydra, every aspect of model training can be modified from the command-line. This is extremely helpful for running lots of experiments on compute clusters or for quickly testing parameters while developing.

All NeMo examples come with instructions on how to run the training/inference script from the command-line (see here for an example).

With Hydra, arguments are set using the = operator:

python examples/asr/asr_ctc/ \
    model.train_ds.manifest_filepath=/path/to/my/train/manifest.json \
    model.validation_ds.manifest_filepath=/path/to/my/validation/manifest.json \
    trainer.devices=2 \
    trainer.accelerator='gpu' \

We can use the + operator to add arguments from the CLI:

python examples/asr/asr_ctc/ \
    model.train_ds.manifest_filepath=/path/to/my/train/manifest.json \
    model.validation_ds.manifest_filepath=/path/to/my/validation/manifest.json \
    trainer.devices=2 \
    trainer.accelerator='gpu' \
    trainer.max_epochs=50 \

We can use the ~ operator to remove configurations:

python examples/asr/asr_ctc/ \
    model.train_ds.manifest_filepath=/path/to/my/train/manifest.json \
    model.validation_ds.manifest_filepath=/path/to/my/validation/manifest.json \
    ~model.test_ds \
    trainer.devices=2 \
    trainer.accelerator='gpu' \
    trainer.max_epochs=50 \

We can specify configuration files using the --config-path and --config-name flags:

python examples/asr/asr_ctc/ \
    --config-path=conf/quartznet \
    --config-name=quartznet_15x5 \
    model.train_ds.manifest_filepath=/path/to/my/train/manifest.json \
    model.validation_ds.manifest_filepath=/path/to/my/validation/manifest.json \
    ~model.test_ds \
    trainer.devices=2 \
    trainer.accelerator='gpu' \
    trainer.max_epochs=50 \


Dataclasses allow NeMo to ship model configurations as part of the NeMo library and also enables pure Python configuration of NeMo models. With Hydra, dataclasses can be used to create structured configs for the conversational AI application.

As an example, refer to the code block below for an Attenion is All You Need machine translation model. The model configuration can be instantiated and modified like any Python Dataclass.

from nemo.collections.nlp.models.machine_translation.mt_enc_dec_config import AAYNBaseConfig

cfg = AAYNBaseConfig()

# modify the number of layers in the encoder
cfg.encoder.num_layers = 8

# modify the training batch size
cfg.train_ds.tokens_in_batch = 8192


Configuration with Hydra always has the following precedence CLI > YAML > Dataclass


Optimizers and learning rate schedules are configurable across all NeMo models and have their own namespace. Here is a sample YAML configuration for a Novograd optimizer with Cosine Annealing learning rate schedule.

    name: novograd
    lr: 0.01

    # optimizer arguments
    betas: [0.8, 0.25]
    weight_decay: 0.001

    # scheduler setup
        name: CosineAnnealing

        # Optional arguments
        max_steps: -1 # computed at runtime or explicitly set here
        monitor: val_loss
        reduce_on_plateau: false

        # scheduler config override
        warmup_steps: 1000
        warmup_ratio: null
        min_lr: 1e-9:


NeMo Examples has optimizer and scheduler configurations for every NeMo model.

Optimizers can be configured from the CLI as well:

python examples/asr/asr_ctc/ \
    --config-path=conf/quartznet \
    --config-name=quartznet_15x5 \
    # train with the adam optimizer
    model.optim=adam \
    # change the learning rate \
    # modify betas
    model.optim.betas=[.8, .5]


name corresponds to the lowercase name of the optimizer. To view a list of available optimizers, run:

from nemo.core.optim.optimizers import AVAILABLE_OPTIMIZERS

for name, opt in AVAILABLE_OPTIMIZERS.items():
    print(f'name: {name}, opt: {opt}')
name: sgd opt: <class 'torch.optim.sgd.SGD'>
name: adam opt: <class 'torch.optim.adam.Adam'>
name: adamw opt: <class 'torch.optim.adamw.AdamW'>
name: adadelta opt: <class 'torch.optim.adadelta.Adadelta'>
name: adamax opt: <class 'torch.optim.adamax.Adamax'>
name: adagrad opt: <class 'torch.optim.adagrad.Adagrad'>
name: rmsprop opt: <class 'torch.optim.rmsprop.RMSprop'>
name: rprop opt: <class 'torch.optim.rprop.Rprop'>
name: novograd opt: <class 'nemo.core.optim.novograd.Novograd'>

Optimizer Params#

Optimizer params can vary between optimizers but the lr param is required for all optimizers. To see the available params for an optimizer, we can look at its corresponding dataclass.

from nemo.core.config.optimizers import NovogradParams

NovogradParams(lr='???', betas=(0.95, 0.98), eps=1e-08, weight_decay=0, grad_averaging=False, amsgrad=False, luc=False, luc_trust=0.001, luc_eps=1e-08)

'???' indicates that the lr argument is required.

Register Optimizer#

To register a new optimizer to be used with NeMo, run:

nemo.core.optim.optimizers.register_optimizer(name: str, optimizer: torch.optim.optimizer.Optimizer, optimizer_params: OptimizerParams)[source]#

Checks if the optimizer name exists in the registry, and if it doesnt, adds it.

This allows custom optimizers to be added and called by name during instantiation.

  • name – Name of the optimizer. Will be used as key to retrieve the optimizer.

  • optimizer – Optimizer class

  • optimizer_params – The parameters as a dataclass of the optimizer

Learning Rate Schedulers#

Learning rate schedulers can be optionally configured under the optim.sched namespace.

name corresponds to the name of the learning rate schedule. To view a list of available schedulers, run:

from nemo.core.optim.lr_scheduler import AVAILABLE_SCHEDULERS

for name, opt in AVAILABLE_SCHEDULERS.items():
    print(f'name: {name}, schedule: {opt}')
name: WarmupPolicy, schedule: <class 'nemo.core.optim.lr_scheduler.WarmupPolicy'>
name: WarmupHoldPolicy, schedule: <class 'nemo.core.optim.lr_scheduler.WarmupHoldPolicy'>
name: SquareAnnealing, schedule: <class 'nemo.core.optim.lr_scheduler.SquareAnnealing'>
name: CosineAnnealing, schedule: <class 'nemo.core.optim.lr_scheduler.CosineAnnealing'>
name: NoamAnnealing, schedule: <class 'nemo.core.optim.lr_scheduler.NoamAnnealing'>
name: WarmupAnnealing, schedule: <class 'nemo.core.optim.lr_scheduler.WarmupAnnealing'>
name: InverseSquareRootAnnealing, schedule: <class 'nemo.core.optim.lr_scheduler.InverseSquareRootAnnealing'>
name: SquareRootAnnealing, schedule: <class 'nemo.core.optim.lr_scheduler.SquareRootAnnealing'>
name: PolynomialDecayAnnealing, schedule: <class 'nemo.core.optim.lr_scheduler.PolynomialDecayAnnealing'>
name: PolynomialHoldDecayAnnealing, schedule: <class 'nemo.core.optim.lr_scheduler.PolynomialHoldDecayAnnealing'>
name: StepLR, schedule: <class 'torch.optim.lr_scheduler.StepLR'>
name: ExponentialLR, schedule: <class 'torch.optim.lr_scheduler.ExponentialLR'>
name: ReduceLROnPlateau, schedule: <class 'torch.optim.lr_scheduler.ReduceLROnPlateau'>
name: CyclicLR, schedule: <class 'torch.optim.lr_scheduler.CyclicLR'>

Scheduler Params#

To see the available params for a scheduler, we can look at its corresponding dataclass:

from nemo.core.config.schedulers import CosineAnnealingParams

CosineAnnealingParams(last_epoch=-1, warmup_steps=None, warmup_ratio=None, min_lr=0.0)

Register scheduler#

To register a new scheduler to be used with NeMo, run:

nemo.core.optim.lr_scheduler.register_scheduler(name: str, scheduler: torch.optim.lr_scheduler._LRScheduler, scheduler_params: SchedulerParams)[source]#

Checks if the scheduler name exists in the registry, and if it doesnt, adds it.

This allows custom schedulers to be added and called by name during instantiation.

  • name – Name of the optimizer. Will be used as key to retrieve the optimizer.

  • scheduler – Scheduler class (inherits from _LRScheduler)

  • scheduler_params – The parameters as a dataclass of the scheduler

Save and Restore#

NeMo models all come with .save_to and .restore_from methods.


To save a NeMo model, run:


Everything needed to use the trained model is packaged and saved in the .nemo file. For example, in the NLP domain, .nemo files include the necessary tokenizer models and/or vocabulary files, etc.


A .nemo file is simply an archive like any other .tar file.


To restore a NeMo model, run:

# Here, you should usually use the class of the model, or simply use ModelPT.restore_from() for simplicity.

When using the PyTorch Lightning Trainer, a PyTorch Lightning checkpoint is created. These are mainly used within NeMo to auto-resume training. Since NeMo models are LightningModules, the PyTorch Lightning method load_from_checkpoint is available. Note that load_from_checkpoint won’t necessarily work out-of-the-box for all models as some models require more artifacts than just the checkpoint to be restored. For these models, the user will have to override load_from_checkpoint if they want to use it.

It’s highly recommended to use restore_from to load NeMo models.

Restore with Modified Config#

Sometimes, there may be a need to modify the model (or it’s sub-components) prior to restoring a model. A common case is when the model’s internal config must be updated due to various reasons (such as deprecation, newer versioning, support a new feature). As long as the model has the same parameters as compared to the original config, the parameters can once again be restored safely.

In NeMo, as part of the .nemo file, the model’s internal config will be preserved. This config is used during restoration, and as shown below we can update this config prior to restoring the model.

# When restoring a model, you should generally use the class of the model
# Obtain the config (as an OmegaConf object)
config = model_class.restore_from('/path/to/model.nemo', return_config=True)
# OR
config = model_class.from_pretrained('name_of_the_model', return_config=True)

# Modify the config as needed
config.x.y = z

# Restore the model from the updated config
model = model_class.restore_from('/path/to/model.nemo', override_config_path=config)
# OR
model = model_class.from_pretrained('name_of_the_model', override_config_path=config)

Register Artifacts#

Conversational AI models can be complicated to restore as more information is needed than just the checkpoint weights in order to use the model. NeMo models can save additional artifacts in the .nemo file by calling .register_artifact. When restoring NeMo models using .restore_from or .from_pretrained, any artifacts that were registered will be available automatically.

As an example, consider an NLP model that requires a trained tokenizer model. The tokenizer model file can be automatically added to the .nemo file with the following:

self.encoder_tokenizer = get_nmt_tokenizer(

By default, .register_artifact will always return a path. If the model is being restored from a .nemo file, then that path will be to the artifact in the .nemo file. Otherwise, .register_artifact will return the local path specified by the user.

config_path is the artifact key. It usually corresponds to a model configuration but does not have to. The model config that is packaged with the .nemo file will be updated according to the config_path key. In the above example, the model config will have

    tokenizer_model: nemo:4978b28103264263a03439aaa6560e5e_tokenizer.model

src is the path to the artifact and the base-name of the path will be used when packaging the artifact in the .nemo file. Each artifact will have a hash prepended to the basename of src in the .nemo file. This is to prevent collisions with basenames base-names that are identical (say when there are two or more tokenizers, both called tokenizer.model). The resulting .nemo file will then have the following file:


If verify_src_exists is set to False, then the artifact is optional. This means that .register_artifact will return None if the src cannot be found.

Nested NeMo Models#

In some cases, it may be helpful to use NeMo models inside other NeMo models. For example, we can incorporate language models into ASR models to use in a decoding process to improve accuracy or use hybrid ASR-TTS models to generate audio from the text on the fly to train or finetune the ASR model.

There are 3 ways to instantiate child models inside parent models:

  • use subconfig directly

  • use the .nemo checkpoint path to load the child model

  • use a pretrained NeMo model

To register a child model, use the register_nemo_submodule method of the parent model. This method will add the child model to a provided model attribute and, in the serialization process, will handle child artifacts correctly and store the child model config in the parent model config in config_field.

from nemo.core.classes import ModelPT

class ChildModel(ModelPT):
    ...  # implement necessary methods

class ParentModel(ModelPT):
    def __init__(self, cfg, trainer=None):
        super().__init__(cfg=cfg, trainer=trainer)

        # optionally annotate type for IDE autocompletion and type checking
        self.child_model: Optional[ChildModel]
        if cfg.get("child_model") is not None:
            # load directly from config
            # either if config provided initially, or automatically
            # after model restoration
                model=ChildModel(self.cfg.child_model, trainer=trainer),
        elif cfg.get('child_model_path') is not None:
            # load from .nemo model checkpoint
            # while saving, config will be automatically assigned/updated
            # in cfg.child_model
                model=ChildModel.restore_from(self.cfg.child_model_path, trainer=trainer),
        elif cfg.get('child_model_name') is not None:
            # load from pretrained model
            # while saving, config will be automatically assigned/updated
            # in cfg.child_model
                model=ChildModel.from_pretrained(self.cfg.child_model_name, trainer=trainer),
            self.child_model = None

Neural Modules#

NeMo is built around Neural Modules, conceptual blocks of neural networks that take typed inputs and produce typed outputs. Such modules typically represent data layers, encoders, decoders, language models, loss functions, or methods of combining activations. NeMo makes it easy to combine and re-use these building blocks while providing a level of semantic correctness checking via its neural type system.


All Neural Modules inherit from ``torch.nn.Module`` and are therefore compatible with the PyTorch ecosystem.

There are 3 types on Neural Modules:

  • Regular modules

  • Dataset/IterableDataset

  • Losses

Every Neural Module in NeMo must inherit from nemo.core.classes.module.NeuralModule class.

class nemo.core.classes.module.NeuralModule(*args: Any, **kwargs: Any)[source]#

Bases: Module, Typing, Serialization, FileIO

Abstract class offering interface shared between all PyTorch Neural Modules.


Context manager which temporarily freezes a module, yields control and finally unfreezes the module.

freeze() None[source]#

Freeze all params for inference.

input_example(max_batch=None, max_dim=None)[source]#

Override this method if random inputs won’t work :returns: A tuple sample of valid input data.

property num_weights#

Utility property that returns the total number of parameters of NeuralModule.

unfreeze() None[source]#

Unfreeze all parameters for training.

Every Neural Modules inherits the nemo.core.classes.common.Typing interface and needs to define neural types for its inputs and outputs. This is done by defining two properties: input_types and output_types. Each property should return an ordered dictionary of “port name”->”port neural type” pairs. Here is the example from ConvASREncoder class:

def input_types(self):
    return OrderedDict(
            "audio_signal": NeuralType(('B', 'D', 'T'), SpectrogramType()),
            "length": NeuralType(tuple('B'), LengthsType()),

def output_types(self):
    return OrderedDict(
            "outputs": NeuralType(('B', 'D', 'T'), AcousticEncodedRepresentation()),
            "encoded_lengths": NeuralType(tuple('B'), LengthsType()),

def forward(self, audio_signal, length=None):
The code snippet above means that nemo.collections.asr.modules.conv_asr.ConvASREncoder expects two arguments:
  • First one, named audio_signal of shape [batch, dimension, time] with elements representing spectrogram values.

  • Second one, named length of shape [batch] with elements representing lengths of corresponding signals.

It also means that .forward(...) and __call__(...) methods each produce two outputs:
  • First one, of shape [batch, dimension, time] but with elements representing encoded representation (AcousticEncodedRepresentation class).

  • Second one, of shape [batch], corresponding to their lengths.


It is a good practice to define types and add @typecheck() decorator to your .forward() method after your module is ready for use by others.


The outputs of .forward(...) method will always be of type torch.Tensor or container of tensors and will work with any other Pytorch code. The type information is attached to every output tensor. If tensors without types is passed to your module, it will not fail, however the types will not be checked. Thus, it is recommended to define input/output types for all your modules, starting with data layers and add @typecheck() decorator to them.


To temporarily disable typechecking, you can enclose your code in `with typecheck.disable_checks():` statement.

Dynamic Layer Freezing#

You can selectively freeze any modules inside a Nemo model by specifying a freezing schedule in the config yaml. Freezing stops any gradient updates to that module, so that its weights are not changed for that step. This can be useful for combatting catastrophic forgetting, for example when finetuning a large pretrained model on a small dataset.

The default approach is to freeze a module for the first N training steps, but you can also enable freezing for a specific range of steps, for example, from step 20 - 100, or even activate freezing from some N until the end of training. You can also freeze a module for the entire training run. Dynamic freezing is specified in training steps, not epochs.

To enable freezing, add the following to your config:

    enabled: true  # set to false if you want to disable freezing

    modules:   # list all of the modules you want to have freezing logic for
      encoder: 200       # module will be frozen for the first 200 training steps
      decoder: [50, -1]  # module will be frozen at step 50 and will remain frozen until training ends
      joint: [10, 100]   # module will be frozen between step 10 and step 100 (step >= 10 and step <= 100)
      transcoder: -1     # module will be frozen for the entire training run