mipcandy.presets.segmentation

Module Contents

Classes

DeepSupervisionWrapper
SegmentationTrainer

API

class mipcandy.presets.segmentation.DeepSupervisionWrapper(loss: torch.nn.Module, *, weight_factors: Sequence[float] | None = None)

Bases: mipcandy.common.Loss

forward(outputs: Sequence[torch.Tensor], targets: Sequence[torch.Tensor]) → tuple[torch.Tensor, dict[str, float]]

class mipcandy.presets.segmentation.SegmentationTrainer(trainer_folder: str | os.PathLike[str], dataloader: torch.utils.data.DataLoader[tuple[torch.Tensor, torch.Tensor]], validation_dataloader: torch.utils.data.DataLoader[tuple[torch.Tensor, torch.Tensor]], *, recoverable: bool = True, profiler: bool = False, device: torch.device | str = 'cpu', console: rich.console.Console = Console())

Bases: mipcandy.training.Trainer

Initialization

num_classes: int = 1

include_background: bool = True

deep_supervision: bool = False

deep_supervision_scales: Sequence[float] | None = None

deep_supervision_weights: Sequence[float] | None = None

_save_preview(x: torch.Tensor, title: str, quality: float, *, is_label: bool = False) → None

apply_non_linearity(x: torch.Tensor, channel_dim: int) → torch.Tensor

save_preview(image: torch.Tensor, label: torch.Tensor, output: torch.Tensor, *, quality: float = 0.75) → None

build_ema(model: torch.nn.Module) → torch.nn.Module

build_criterion() → torch.nn.Module

build_optimizer(params: mipcandy.types.Params) → torch.optim.Optimizer

build_scheduler(optimizer: torch.optim.Optimizer, num_epochs: int) → torch.optim.lr_scheduler.LRScheduler

backward(images: torch.Tensor, labels: torch.Tensor, toolbox: mipcandy.training.TrainerToolbox) → tuple[float, dict[str, float]]

static prepare_deep_supervision_targets(labels: torch.Tensor, output_shapes: list[tuple[int, ...]]) → list[torch.Tensor]

class_percentages(ids: torch.Tensor) → dict[int, float]

static format_class_percentages(percentages: dict[int, float], prefix: str) → dict[str, float]

validate_case(idx: int, image: torch.Tensor, label: torch.Tensor, toolbox: mipcandy.training.TrainerToolbox) → tuple[float, dict[str, float], torch.Tensor]