pathforge.training

Trainer abstractions and the PyTorch Lightning implementation. Trainers are registered in TRAINERS and resolved by name at runtime.

Base Abstractions

class pathforge.training.base.TrainerBase[source]

Bases: ABC

Framework agnostic trainer base class.

abstractmethod fit(model: MILModelBase, dataset_train: DatasetBase, dataset_val: DatasetBase, loss_func: BaseLoss) → Any[source]

Train a model on train/validation datasets and return trainer-specific results.

Parameters:
Return type:

Any

abstractmethod predict(model: MILModelBase, dataset: DatasetBase) → Any[source]

Run inference on a dataset using a trained model.

Parameters:
Return type:

Any

class pathforge.training.base.MILTrainer[source]

Bases: object

Convenience façade specifically for MIL use cases.

This constrains the types to MILModelBase + BagDatasetBase, which matches the architecture diagram’s “MILTrainer” box.

trainer: TrainerBase
model: MILModelBase
dataset_train: BagDatasetBase
dataset_val: BagDatasetBase
loss: BaseLoss
run() → Any[source]
Return type:

Any

__init__(trainer: TrainerBase, model: MILModelBase, dataset_train: BagDatasetBase, dataset_val: BagDatasetBase, loss: BaseLoss) → None
Parameters:
Return type:

None

Lightning Trainer

The Lightning trainer is registered as "lightning" in TRAINERS.

class pathforge.training.lightning.LightningModuleAdapter[source]

Bases: LightningModule

Adapter: Wraps a PathForge MILModelBase into a PL LightningModule. Handles optimization logic, logging, and scheduler configuration via Config. Pass loss_fn=None only for inference (predict_step does not use it).

__init__(model: MILModelBase, loss_fn: BaseLoss | None, config: Config)[source]
Parameters:
forward(*args, **kwargs)[source]

Same as torch.nn.Module.forward().

Parameters:

  • *args – Whatever you decide to pass into the forward method.

  • **kwargs – Keyword arguments are also possible.

Returns:

Your model’s output

training_step(batch: Any, batch_idx: int) → Tensor[source]

Here you compute and return the training loss and some additional metrics for e.g. the progress bar or logger.

Parameters:

  • batch (Any) – The output of your data iterable, normally a DataLoader.

  • batch_idx (int) – The index of this batch.

  • dataloader_idx – The index of the dataloader that produced this batch. (only if multiple dataloaders used)

Returns:

  • Tensor - The loss tensor

  • dict - A dictionary which can include any keys, but must include the key 'loss' in the case of automatic optimization.

  • None - In automatic optimization, this will skip to the next batch (but is not supported for multi-GPU, TPU, or DeepSpeed). For manual optimization, this has no special meaning, as returning the loss is not required.

Return type:

Tensor

In this step you’d normally do the forward pass and calculate the loss for a batch. You can also do fancier things like multiple forward passes or something model specific.

Example:

def training_step(self, batch, batch_idx):
    x, y, z = batch
    out = self.encoder(x)
    loss = self.loss(out, x)
    return loss

To use multiple optimizers, you can switch to ‘manual optimization’ and control their stepping:

def __init__(self):
    super().__init__()
    self.automatic_optimization = False


# Multiple optimizers (e.g.: GANs)
def training_step(self, batch, batch_idx):
    opt1, opt2 = self.optimizers()

    # do training_step with encoder
    ...
    opt1.step()
    # do training_step with decoder
    ...
    opt2.step()

Note

When accumulate_grad_batches > 1, the loss returned here will be automatically normalized by accumulate_grad_batches internally.

validation_step(batch: Any, batch_idx: int) → Tensor[source]

Operates on a single batch of data from the validation set. In this step you’d might generate examples or calculate anything of interest like accuracy.

Parameters:

  • batch (Any) – The output of your data iterable, normally a DataLoader.

  • batch_idx (int) – The index of this batch.

  • dataloader_idx – The index of the dataloader that produced this batch. (only if multiple dataloaders used)

Returns:

  • Tensor - The loss tensor

  • dict - A dictionary. Can include any keys, but must include the key 'loss'.

  • None - Skip to the next batch.

Return type:

Tensor

# if you have one val dataloader:
def validation_step(self, batch, batch_idx): ...


# if you have multiple val dataloaders:
def validation_step(self, batch, batch_idx, dataloader_idx=0): ...

Examples:

# CASE 1: A single validation dataset
def validation_step(self, batch, batch_idx):
    x, y = batch

    # implement your own
    out = self(x)
    loss = self.loss(out, y)

    # log 6 example images
    # or generated text... or whatever
    sample_imgs = x[:6]
    grid = torchvision.utils.make_grid(sample_imgs)
    self.logger.experiment.add_image('example_images', grid, 0)

    # calculate acc
    labels_hat = torch.argmax(out, dim=1)
    val_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0)

    # log the outputs!
    self.log_dict({'val_loss': loss, 'val_acc': val_acc})

If you pass in multiple val dataloaders, validation_step() will have an additional argument. We recommend setting the default value of 0 so that you can quickly switch between single and multiple dataloaders.

# CASE 2: multiple validation dataloaders
def validation_step(self, batch, batch_idx, dataloader_idx=0):
    # dataloader_idx tells you which dataset this is.
    x, y = batch

    # implement your own
    out = self(x)

    if dataloader_idx == 0:
        loss = self.loss0(out, y)
    else:
        loss = self.loss1(out, y)

    # calculate acc
    labels_hat = torch.argmax(out, dim=1)
    acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0)

    # log the outputs separately for each dataloader
    self.log_dict({f"val_loss_{dataloader_idx}": loss, f"val_acc_{dataloader_idx}": acc})

Note

If you don’t need to validate you don’t need to implement this method.

Note

When the validation_step() is called, the model has been put in eval mode and PyTorch gradients have been disabled. At the end of validation, the model goes back to training mode and gradients are enabled.

predict_step(batch: Any, batch_idx: int, dataloader_idx: int = 0) → Tensor[source]

Step function called during predict(). By default, it calls forward(). Override to add any processing logic.

The predict_step() is used to scale inference on multi-devices.

To prevent an OOM error, it is possible to use BasePredictionWriter callback to write the predictions to disk or database after each batch or on epoch end.

The BasePredictionWriter should be used while using a spawn based accelerator. This happens for Trainer(strategy="ddp_spawn") or training on 8 TPU cores with Trainer(accelerator="tpu", devices=8) as predictions won’t be returned.

Parameters:

  • batch (Any) – The output of your data iterable, normally a DataLoader.

  • batch_idx (int) – The index of this batch.

  • dataloader_idx (int) – The index of the dataloader that produced this batch. (only if multiple dataloaders used)

Returns:

Predicted output (optional).

Return type:

Tensor

Example

class MyModel(LightningModule):

    def predict_step(self, batch, batch_idx, dataloader_idx=0):
        return self(batch)

dm = ...
model = MyModel()
trainer = Trainer(accelerator="gpu", devices=2)
predictions = trainer.predict(model, dm)
on_validation_epoch_start() → None[source]

Called in the validation loop at the very beginning of the epoch.

Return type:

None

on_validation_epoch_end() → None[source]

Called in the validation loop at the very end of the epoch.

Return type:

None

configure_optimizers()[source]

Build the configured optimizer and optional learning-rate scheduler.

class pathforge.training.lightning.LightningTrainer[source]

Bases: TrainerBase

Trainer implementation using PyTorch Lightning.

Config Usage: - mil.epochs -> max_epochs - mil.accumulate_grad_batches - mil.gradient_clip_val - experiment.num_workers

__init__(config: Config, extra_callbacks: List[Callback] | None = None)[source]
Parameters:

  • config (Config)

  • extra_callbacks (List[Callback] | None)

fit(model: MILModelBase, dataset_train: Dataset, dataset_val: Dataset, loss_func: BaseLoss) → Tuple[str, float][source]

Train the model. :returns: (best_model_path, best_model_score)

Parameters:
Return type:

Tuple[str, float]

predict(model: MILModelBase, dataset: Dataset) → Tensor[source]

Run inference. loss_fn is not used during prediction.

Parameters:
Return type:

Tensor

Callbacks

Metrics

class pathforge.training.metrics.TaskEvaluationArtifacts[source]

Bases: object

Persisted evaluation outputs for one trained model.

Variables:

  • metrics_path (pathlib.Path) – JSON file containing scalar task metrics.

  • curve_data_path (pathlib.Path) – JSON file containing inspectable curve coordinates.

  • figure_paths (dict[str, pathlib.Path]) – Mapping from figure name to the saved PNG path.

metrics_path: Path
curve_data_path: Path
figure_paths: dict[str, Path]
__init__(metrics_path: Path, curve_data_path: Path, figure_paths: dict[str, Path]) → None
Parameters:

  • metrics_path (Path)

  • curve_data_path (Path)

  • figure_paths (dict[str, Path])

Return type:

None

pathforge.training.metrics.compute_task_metrics(predictions: Tensor, target: Any, *, task: str, selected_metrics: list[str] | None = None) → dict[str, float][source]

Compute scalar validation metrics for one PathForge task.

Parameters:

  • predictions (Tensor) – Model output tensor. Classification expects logits shaped [B] or [B, C]. Continuous survival expects risk/log-hazard shaped [B] or [B, 1]. Discrete survival expects logits shaped [B, T].

  • target (Any) – Task target. Classification expects integer labels shaped [B]. Survival tasks expect a dictionary with time and event tensors shaped [B].

  • task (str) – PathForge task name.

  • selected_metrics (list[str] | None)

Returns:

Finite-or-NaN scalar metrics ready for Lightning logging and JSON serialization.

Return type:

dict[str, float]

pathforge.training.metrics.save_task_evaluation_artifacts(predictions: Tensor, target: Any, *, task: str, output_dir: Path, prefix: str, selected_metrics: list[str] | None = None) → TaskEvaluationArtifacts[source]

Write scalar metrics plus task-specific visualization artifacts.

Parameters:

  • predictions (Tensor) – Task prediction tensor with the same shape contract as compute_task_metrics().

  • target (Any) – Task target object with the same contract as compute_task_metrics().

  • task (str) – PathForge task name.

  • output_dir (Path) – Directory receiving JSON and PNG artifacts.

  • prefix (str) – Output prefix such as "val".

  • selected_metrics (list[str] | None)

Returns:

Paths to the persisted JSON and PNG files.

Return type:

TaskEvaluationArtifacts

Scikit-Learn Trainer

Trainer for scikit-learn / scikit-survival slide-level estimators.

Usage:

from pathforge.core.models.sklearn_slide import SklearnLogisticRegressionClassifier
from pathforge.training.sklearn_trainer import SklearnSlideTrainer

model = SklearnLogisticRegressionClassifier()
trainer = SklearnSlideTrainer(output_dir=Path("runs/lr"), task="classification")
model_path, score = trainer.fit(model, X_train, y_train, X_val, y_val)
class pathforge.training.sklearn_trainer.SklearnSlideTrainer[source]

Bases: TrainerBase

Fit a ScikitBase estimator on slide-level numpy feature arrays and persist evaluation artefacts.

This trainer does not use PyTorch Lightning; it calls sklearn’s fit/predict directly and then runs the same save_task_evaluation_artifacts() pipeline used by LightningTrainer.

Parameters:

  • output_dir – Root directory; model pickle and artifacts go here.

  • task – PathForge task name — one of "classification", "regression", "survival", "survival_discrete".

  • selected_metrics – Optional list of metric names to compute; defaults to all metrics for the task.

__init__(output_dir: Path | str, task: str, selected_metrics: list[str] | None = None) → None[source]
Parameters:

  • output_dir (Path | str)

  • task (str)

  • selected_metrics (list[str] | None)

Return type:

None

fit(model: ScikitBase, X_train: ndarray, y_train: Any, X_val: ndarray, y_val: Any) → tuple[str, float][source]

Fit model and evaluate on the validation set.

Parameters:

  • model (ScikitBase) – Unfitted sklearn estimator wrapper.

  • X_train (ndarray) – Training feature matrix shaped [N_train, D].

  • y_train (Any) – Training targets. For classification/regression: a 1-D numpy array; for survival: a dict with "time" and "event" keys (1-D numpy arrays).

  • X_val (ndarray) – Validation feature matrix shaped [N_val, D].

  • y_val (Any) – Validation targets with the same structure as y_train.

Returns:

Tuple of (model_pickle_path, best_scalar_score) where the score is the primary validation metric for the task (balanced_accuracy for classification, c_index for survival, r2 for regression).

Return type:

tuple[str, float]

predict(model: ScikitBase, dataset: ndarray) → Tensor[source]

Run prediction for a slide-level feature matrix shaped [N, D].

Parameters:
Return type:

Tensor