"""PakFlood task: the authors' FloodSegmentationTask plus - loss "cedice": class-weighted cross-entropy + Dice (on chosen classes) - init_from_checkpoint: load a weights-only checkpoint before training (few-shot pass from the hold-out model) """ from typing import Any import segmentation_models_pytorch as smp import torch from torch import nn from geoid_flood.tasks.flood_segmentation import FloodSegmentationTask class CEDiceLoss(nn.Module): """class-weighted CE + dice_weight * multiclass Dice (ignore_index aware).""" def __init__(self, class_weights, ignore_index: int, dice_weight: float = 1.0, dice_classes=None): super().__init__() w = torch.tensor(class_weights, dtype=torch.float32) if class_weights is not None else None self.ce = nn.CrossEntropyLoss(weight=w, ignore_index=ignore_index) self.dice = smp.losses.DiceLoss("multiclass", classes=dice_classes, ignore_index=ignore_index) self.dice_weight = dice_weight def forward(self, y_hat: torch.Tensor, y: torch.Tensor) -> torch.Tensor: return self.ce(y_hat, y) + self.dice_weight * self.dice(y_hat, y) class FloodCEDiceTask(FloodSegmentationTask): """Extra init args (all optional): dice_weight (float), dice_classes (list[int], e.g. [2] = flood only), init_from_checkpoint (path to *_state_dict.ckpt). With loss != "cedice" it behaves exactly like FloodSegmentationTask. """ def __init__( self, *args: Any, dice_weight: float = 1.0, dice_classes: list[int] | None = None, init_from_checkpoint: str | None = None, **kwargs: Any, ): self.dice_weight = dice_weight self.dice_classes = dice_classes super().__init__(*args, **kwargs) if init_from_checkpoint: ck = torch.load(init_from_checkpoint, map_location="cpu", weights_only=False) sd = ck.get("state_dict", ck) if isinstance(ck, dict) else ck missing, unexpected = self.load_state_dict(sd, strict=False) n_loaded = len(sd) - len(unexpected) print(f"[pakflood] init_from_checkpoint {init_from_checkpoint}: loaded {n_loaded} tensors, " f"missing {len(missing)}, unexpected {len(unexpected)}", flush=True) if missing: print("[pakflood] first missing keys:", missing[:5], flush=True) if n_loaded < 0.9 * len(self.state_dict()): raise RuntimeError("init_from_checkpoint loaded less than 90% of the model; architecture mismatch?") def configure_losses(self) -> None: if self.hparams["loss"] != "cedice": return super().configure_losses() self.criterion = CEDiceLoss( self.hparams["class_weights"], self.hparams["ignore_index"], dice_weight=self.dice_weight, dice_classes=self.dice_classes, )