Implementation:Snorkel team Snorkel Trainer Fit Multitask

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

Overview

Concrete tool for training MultitaskClassifier models with configurable optimization and monitoring, provided by the Snorkel library. This is the same Trainer class used for slice-aware training, documented here for the general multi-task classification context.

Description

The Trainer class provides the training loop for MultitaskClassifier. For the general multi-task context, the key difference from slice-aware usage is that tasks may have independent dataloaders and different label spaces.

Key features for multi-task training:

• Shuffled batch scheduler: Interleaves batches from all tasks randomly
• Sequential batch scheduler: Processes one task at a time
• Per-task loss computation: Each task uses its configured loss function
• Shared optimizer: Single optimizer updates all shared and task-specific parameters

Usage

Import this class when training a MultitaskClassifier. Same class as used for slice-aware training.

Code Reference

Source Location

• Repository: snorkel
• File: snorkel/classification/training/trainer.py
• Lines: L108-586 (Trainer class), L35-105 (TrainerConfig)

Signature

class Trainer:
    def __init__(self, name: Optional[str] = None, **kwargs: Any) -> None:
        """
        Args:
            name: Trainer name.
            **kwargs: TrainerConfig overrides (n_epochs, lr, l2,
                grad_clip, optimizer, lr_scheduler, batch_scheduler,
                checkpointing, logging, log_writer, train_split,
                valid_split, test_split).
        """

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

        Args:
            model: MultitaskClassifier to train.
            dataloaders: DictDataLoaders for 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
dataloaders	List[DictDataLoader]	Yes	Dataloaders with named splits
n_epochs	int	No	Training epochs (default 1)
lr	float	No	Learning rate (default 0.01)
batch_scheduler	str	No	"shuffled" or "sequential" (default "shuffled")
checkpointing	bool	No	Save best model (default False)
logging	bool	No	Enable metric logging (default False)

Outputs

Name	Type	Description
Trained model	MultitaskClassifier	Model updated in-place
Checkpoints	Files	Optional saved model states
Logs	Files	Optional JSON or TensorBoard logs

Usage Examples

from snorkel.classification import Trainer

trainer = Trainer(
    n_epochs=20,
    lr=0.001,
    optimizer="adam",
    batch_scheduler="shuffled",
    checkpointing=True,
    checkpointer_config={"checkpoint_dir": "./checkpoints"},
    logging=True,
    log_writer="tensorboard",
)

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

Related Pages

Implements Principle

• Principle:Snorkel_team_Snorkel_Multitask_Training_Execution

Requires Environment

• Environment:Snorkel_team_Snorkel_PyTorch