fishy.experiments.transfer¶

Transfer learning module for deep learning models. Standardized to use DataModule and Trainer patterns.

Functions

fishy.experiments.transfer.run_sequential_transfer_learning(model_name: str, transfer_datasets: List[str], target_dataset: str, num_epochs_transfer: int = 10, num_epochs_finetune: int = 20, batch_size: int = 32, learning_rate: float = 0.001, finetune_lr: float = 0.0005, device: str = 'cpu', save_intermediate: bool = False, val_split: float = 0.2, file_path: str | None = None, wandb_project: str = 'fishy-business', wandb_entity: str = 'victoria-university-of-wellington', wandb_log: bool = False, run: int = 0, wandb_run: Any | None = None) → Tuple[Module, Dict[str, Any]][source]¶

Performs sequential transfer learning using standardized DataModules.

Examples

>>> m_name = "transformer"
>>> isinstance(m_name, str)
True

Parameters:

model_name (str) – Name of the model architecture.
transfer_datasets (List[str]) – List of datasets to pre-train on.
target_dataset (str) – Final dataset to fine-tune on.
num_epochs_transfer (int) – Epochs per transfer phase.
num_epochs_finetune (int) – Epochs for final phase.
batch_size (int) – Batch size.
learning_rate (float) – Initial learning rate.
finetune_lr (float) – Learning rate for fine-tuning.
device (str) – Computation device.
save_intermediate (bool) – Save checkpoints after each phase.
val_split (float) – Fraction of data for validation.
file_path (str) – Path to data file.
wandb_project (str) – W&B project name.
wandb_entity (str) – W&B entity.
wandb_log (bool) – Enable W&B logging.
run (int) – Run identifier/seed.
wandb_run (Any) – Existing W&B run.

Returns:

Trained model and history.

Return type:

Tuple[nn.Module, Dict[str, Any]]