sleap.nn.training

Training functionality and high level APIs.

class sleap.nn.training.BottomUpModelTrainer(data_readers: DataReaders, model: Model, config: TrainingJobConfig, initial_config: Optional[TrainingJobConfig] = None)

Trainer for models that output multi-instance confidence maps and PAFs.

property has_offsets: bool

Whether model is configured to output refinement offsets.

property input_keys: List[str]

Return example keys to be mapped to model inputs.

property output_keys: List[str]

Return example keys to be mapped to model outputs.

class sleap.nn.training.BottomUpMultiClassModelTrainer(data_readers: DataReaders, model: Model, config: TrainingJobConfig, initial_config: Optional[TrainingJobConfig] = None)

Trainer for models that output multi-instance confidence maps and class maps.

property has_offsets: bool

Whether model is configured to output refinement offsets.

property input_keys: List[str]

Return example keys to be mapped to model inputs.

property output_keys: List[str]

Return example keys to be mapped to model outputs.

class sleap.nn.training.CentroidConfmapsModelTrainer(data_readers: DataReaders, model: Model, config: TrainingJobConfig, initial_config: Optional[TrainingJobConfig] = None)

Trainer for models that output centroid confidence maps.

property has_offsets: bool

Whether model is configured to output refinement offsets.

property input_keys: List[str]

Return example keys to be mapped to model inputs.

property output_keys: List[str]

Return example keys to be mapped to model outputs.

class sleap.nn.training.DataReaders(training_labels_reader: LabelsReader, validation_labels_reader: LabelsReader, test_labels_reader: Optional[LabelsReader] = None)

Container class for SLEAP labels that serve as training data sources.

training_labels_reader

LabelsReader pipeline provider for a training data from a Labels instance.

Type:

sleap.nn.data.providers.LabelsReader

validation_labels_reader

LabelsReader pipeline provider for a validation data from a Labels instance.

Type:

sleap.nn.data.providers.LabelsReader

test_labels_reader

LabelsReader pipeline provider for a test set data from a Labels instance. This is not necessary for training.

Type:

Optional[sleap.nn.data.providers.LabelsReader]

classmethod from_config(labels_config: LabelsConfig, training: Union[str, Labels], validation: Union[str, Labels, float], test: Optional[Union[str, Labels]] = None, video_search_paths: Optional[List[str]] = None, update_config: bool = False, with_track_only: bool = False) → DataReaders

Create data readers from a (possibly incomplete) configuration.

classmethod from_labels(training: Union[str, Labels], validation: Union[str, Labels, float], test: Optional[Union[str, Labels]] = None, video_search_paths: Optional[List[str]] = None, labels_config: Optional[LabelsConfig] = None, update_config: bool = False, with_track_only: bool = False) → DataReaders

Create data readers from Labels datasets as data providers.

property test_labels: Labels

Return the Labels underlying the test data reader.

property training_labels: Labels

Return the Labels underlying the training data reader.

property validation_labels: Labels

Return the Labels underlying the validation data reader.

class sleap.nn.training.SingleInstanceModelTrainer(data_readers: DataReaders, model: Model, config: TrainingJobConfig, initial_config: Optional[TrainingJobConfig] = None)

Trainer for models that output single-instance confidence maps.

property has_offsets: bool

Whether model is configured to output refinement offsets.

property input_keys: List[str]

Return example keys to be mapped to model inputs.

property output_keys: List[str]

Return example keys to be mapped to model outputs.

class sleap.nn.training.TopDownMultiClassModelTrainer(data_readers: DataReaders, model: Model, config: TrainingJobConfig, initial_config: Optional[TrainingJobConfig] = None)

Trainer for models that output multi-instance confidence maps and class maps.

property has_offsets: bool

Whether model is configured to output refinement offsets.

property input_keys: List[str]

Return example keys to be mapped to model inputs.

property output_keys: List[str]

Return example keys to be mapped to model outputs.

class sleap.nn.training.TopdownConfmapsModelTrainer(data_readers: DataReaders, model: Model, config: TrainingJobConfig, initial_config: Optional[TrainingJobConfig] = None)

Trainer for models that output instance centered confidence maps.

property has_offsets: bool

Whether model is configured to output refinement offsets.

property input_keys: List[str]

Return example keys to be mapped to model inputs.

property output_keys: List[str]

Return example keys to be mapped to model outputs.

class sleap.nn.training.Trainer(data_readers: DataReaders, model: Model, config: TrainingJobConfig, initial_config: Optional[TrainingJobConfig] = None)

Base trainer class that provides general model training functionality.

This class is intended to be instantiated using the from_config() class method, which will return the appropriate subclass based on the input configuration.

This class should not be used directly. It is intended to be subclassed by a model output type-specific trainer that provides more specific functionality.

data_readers

A DataReaders instance that contains training data providers.

Type:

sleap.nn.training.DataReaders

model

A Model instance describing the SLEAP model to train.

Type:

sleap.nn.model.Model

config

A TrainingJobConfig that describes the training parameters.

Type:

sleap.nn.config.training_job.TrainingJobConfig

initial_config

This attribute will contain a copy of the input configuration before any attributes are updated in config.

Type:

Optional[sleap.nn.config.training_job.TrainingJobConfig]

pipeline_builder

A model output type-specific data pipeline builder to create pipelines that generate data used for training. This must be specified in subclasses.

Type:

sleap.nn.training.PipelineBuilder

training_pipeline

The data pipeline that generates examples from the training set for optimization.

Type:

sleap.nn.data.pipelines.Pipeline

validation_pipeline

The data pipeline that generates examples from the validation set for optimization.

Type:

sleap.nn.data.pipelines.Pipeline

training_viz_pipeline

The data pipeline that generates examples from the training set for visualization.

Type:

sleap.nn.data.pipelines.Pipeline

validation_viz_pipeline

The data pipeline that generates examples from the validation set for visualization.

Type:

sleap.nn.data.pipelines.Pipeline

optimization_callbacks

Keras callbacks related to optimization.

Type:

List[keras.callbacks.Callback]

output_callbacks

Keras callbacks related to outputs.

Type:

List[keras.callbacks.Callback]

visualization_callbacks

Keras callbacks related to visualization.

Type:

List[keras.callbacks.Callback]

run_path

The path to the run folder that will contain training results, if any.

Type:

Optional[str]

property callbacks: List[Callback]

Return all callbacks currently configured.

cleanup()

Delete visualization images subdirectory.

evaluate()

Compute evaluation metrics on data splits and save them.

classmethod from_config(config: TrainingJobConfig, training_labels: Optional[Union[str, Labels]] = None, validation_labels: Optional[Union[str, Labels, float]] = None, test_labels: Optional[Union[str, Labels]] = None, video_search_paths: Optional[List[str]] = None) → Trainer

Initialize the trainer from a training job configuration.

Parameters:

• config – A TrainingJobConfig instance.

• training_labels – Training labels to use instead of the ones in the config, if any. If a path is specified, it will overwrite the one in the config.

• validation_labels – Validation labels to use instead of the ones in the config, if any. If a path is specified, it will overwrite the one in the config.

• test_labels – Teset labels to use instead of the ones in the config, if any. If a path is specified, it will overwrite the one in the config.

abstract property input_keys

Implement in subclasses.

property keras_model: Model

Alias for self.model.keras_model.

abstract property output_keys

Implement in subclasses.

package()

Package model folder into a zip file for portability.

setup()

Set up data pipeline and model for training.

train()

Execute the optimization loop to train the model.

sleap.nn.training.create_trainer_using_cli(args: Optional[List] = None)

Create CLI for training.

sleap.nn.training.get_timestamp() → str

Return the date and time as a string.

sleap.nn.training.main(args: Optional[List] = None)

Create CLI for training and run.

sleap.nn.training.sanitize_scope_name(name: str) → str

Sanitizes string which will be used as TensorFlow scope name.

sleap.nn.training.setup_checkpointing(config: CheckpointingConfig, run_path: str) → List[Callback]

Set up model checkpointing callbacks from config.

sleap.nn.training.setup_losses(config: OptimizationConfig) → Callable[[Tensor, Tensor], Tensor]

Set up model loss function from config.

sleap.nn.training.setup_metrics(config: OptimizationConfig, part_names: Optional[List[str]] = None) → List[Union[Loss, Metric]]

Set up training metrics from config.

sleap.nn.training.setup_new_run_folder(config: OutputsConfig, base_run_name: Optional[str] = None) → str

Create a new run folder from config.

sleap.nn.training.setup_optimization_callbacks(config: OptimizationConfig) → List[Callback]

Set up optimization callbacks from config.

sleap.nn.training.setup_optimizer(config: OptimizationConfig) → OptimizerV2

Set up model optimizer from config.

sleap.nn.training.setup_output_callbacks(config: OutputsConfig, run_path: Optional[str] = None) → List[Callback]

Set up training outputs callbacks from config.

sleap.nn.training.setup_tensorboard(config: TensorBoardConfig, run_path: str) → List[Callback]

Set up TensorBoard callbacks from config.

sleap.nn.training.setup_visualization(config: OutputsConfig, run_path: str, viz_fn: Callable[[], Figure], name: str) → List[Callback]

Set up visualization callbacks from config.

sleap.nn.training.setup_zmq_callbacks(zmq_config: ZMQConfig) → List[Callback]

Set up ZeroMQ callbacks from config.