Implementation:Snorkel team Snorkel Trainer Fit

Knowledge Sources	Snorkel
Domains	Training, Multi_Task_Learning, PyTorch
Last Updated	2026-02-14 20:00 GMT

Overview

Concrete tool for training multi-task classifiers with configurable optimization, logging, and checkpointing, provided by the Snorkel library.

Description

The Trainer class provides a complete training loop for MultitaskClassifier and its subclasses (including SliceAwareClassifier). It handles:

• Optimizer creation (SGD, Adam, Adamax)
• Learning rate scheduling (constant, linear, exponential, step)
• Batch scheduling across multiple tasks (shuffled or sequential)
• Gradient clipping
• Optional metric logging (JSON or TensorBoard)
• Optional model checkpointing (saving best model by metric)

Usage

Import this class when training any Snorkel MultitaskClassifier. Configure via keyword arguments that are merged into TrainerConfig.

Code Reference

Source Location

• Repository: snorkel
• File: snorkel/classification/training/trainer.py
• Lines: L108-586 (class), L138-142 (__init__), L144-248 (fit)

Signature

class Trainer:
    def __init__(self, name: Optional[str] = None, **kwargs: Any) -> None:
        """
        Args:
            name: Trainer name.
            **kwargs: Merged into TrainerConfig:
                n_epochs (int, default 1), lr (float, default 0.01),
                l2 (float, default 0.0), grad_clip (float, default 1.0),
                optimizer (str, default "adam"),
                lr_scheduler (str, default "constant"),
                batch_scheduler (str, default "shuffled"),
                checkpointing (bool, default False),
                logging (bool, default False),
                log_writer (str, default "tensorboard").
        """

    def fit(
        self,
        model: MultitaskClassifier,
        dataloaders: List[DictDataLoader],
    ) -> None:
        """
        Train model on provided dataloaders.

        Args:
            model: MultitaskClassifier (or SliceAwareClassifier) to train.
            dataloaders: List of DictDataLoaders with train/valid/test splits.
        """

Import

from snorkel.classification import Trainer

I/O Contract

Inputs

Name	Type	Required	Description
model	MultitaskClassifier	Yes	Model to train (in-place modification)
dataloaders	List[DictDataLoader]	Yes	Train/valid/test dataloaders
n_epochs	int	No	Training epochs (default 1)
lr	float	No	Learning rate (default 0.01)
optimizer	str	No	"sgd", "adam", or "adamax" (default "adam")
checkpointing	bool	No	Save best model checkpoint (default False)

Outputs

Name	Type	Description
Trained model	MultitaskClassifier	Model updated in-place with trained parameters
Checkpoints	Files	Optional model checkpoints saved to disk
Logs	Files	Optional training logs (JSON or TensorBoard)

Usage Examples

from snorkel.classification import Trainer

# Train slice-aware model
trainer = Trainer(
    n_epochs=10,
    lr=0.001,
    l2=0.01,
    optimizer="adam",
    lr_scheduler="step",
    progress_bar=True,
)

trainer.fit(
    model=model,
    dataloaders=[train_dl, valid_dl],
)

Related Pages

Implements Principle

• Principle:Snorkel_team_Snorkel_Multitask_Model_Training

Requires Environment

• Environment:Snorkel_team_Snorkel_PyTorch