Diceloss

pytorch中的 Diceloss 记录

https://github.com/BloodAxe/pytorch-toolbelt/blob/0.4.1/pytorch_toolbelt/losses/dice.py
输入：
y_pred: NxCxHxW y_true: NxHxW

from typing import List

import torch
import torch.nn.functional as F
from pytorch_toolbelt.utils.torch_utils import to_tensor
from torch import Tensor
from torch.nn.modules.loss import _Loss

from .functional import soft_dice_score

__all__ = ["DiceLoss"]

BINARY_MODE = "binary"
MULTICLASS_MODE = "multiclass"
MULTILABEL_MODE = "multilabel"


class DiceLoss(_Loss):
    """
    Implementation of Dice loss for image segmentation task.
    It supports binary, multiclass and multilabel cases
    """

    def __init__(
        self,
        mode: str,
        classes: List[int] = None,
        log_loss=False,
        from_logits=True,
        smooth: float = 0.0,
        ignore_index=None,
        eps=1e-7,
    ):
        """
        :param mode: Metric mode {'binary', 'multiclass', 'multilabel'}
        :param classes: Optional list of classes that contribute in loss computation;
        By default, all channels are included.
        :param log_loss: If True, loss computed as `-log(jaccard)`; otherwise `1 - jaccard`
        :param from_logits: If True assumes input is raw logits
        :param smooth:
        :param ignore_index: Label that indicates ignored pixels (does not contribute to loss)
        :param eps: Small epsilon for numerical stability
        """
        assert mode in {BINARY_MODE, MULTILABEL_MODE, MULTICLASS_MODE}
        super(DiceLoss, self).__init__()
        self.mode = mode
        if classes is not None:
            assert mode != BINARY_MODE, "Masking classes is not supported with mode=binary"
            classes = to_tensor(classes, dtype=torch.long)

        self.classes = classes
        self.from_logits = from_logits
        self.smooth = smooth
        self.eps = eps
        self.ignore_index = ignore_index
        self.log_loss = log_loss

    def forward(self, y_pred: Tensor, y_true: Tensor) -> Tensor:
        """
        :param y_pred: NxCxHxW
        :param y_true: NxHxW
        :return: scalar
        """
        assert y_true.size(0) == y_pred.size(0)

        if self.from_logits:
            # Apply activations to get [0..1] class probabilities
            # Using Log-Exp as this gives more numerically stable result and does not cause vanishing gradient on
            # extreme values 0 and 1
            if self.mode == MULTICLASS_MODE:
                y_pred = y_pred.log_softmax(dim=1).exp()
            else:
                y_pred = F.logsigmoid(y_pred).exp()

        bs = y_true.size(0)
        num_classes = y_pred.size(1)
        dims = (0, 2)

        if self.mode == BINARY_MODE:
            y_true = y_true.view(bs, 1, -1)
            y_pred = y_pred.view(bs, 1, -1)

            if self.ignore_index is not None:
                mask = y_true != self.ignore_index
                y_pred = y_pred * mask
                y_true = y_true * mask

        if self.mode == MULTICLASS_MODE:
            y_true = y_true.view(bs, -1)
            y_pred = y_pred.view(bs, num_classes, -1)

            if self.ignore_index is not None:
                mask = y_true != self.ignore_index
                y_pred = y_pred * mask.unsqueeze(1)

                y_true = F.one_hot((y_true * mask).to(torch.long), num_classes)  # N,H*W -> N,H*W, C
                y_true = y_true.permute(0, 2, 1) * mask.unsqueeze(1)  # H, C, H*W
            else:
                y_true = F.one_hot(y_true, num_classes)  # N,H*W -> N,H*W, C
                y_true = y_true.permute(0, 2, 1)  # H, C, H*W

        if self.mode == MULTILABEL_MODE:
            y_true = y_true.view(bs, num_classes, -1)
            y_pred = y_pred.view(bs, num_classes, -1)

            if self.ignore_index is not None:
                mask = y_true != self.ignore_index
                y_pred = y_pred * mask
                y_true = y_true * mask

        scores = soft_dice_score(y_pred, y_true.type_as(y_pred), smooth=self.smooth, eps=self.eps, dims=dims)

        if self.log_loss:
            loss = -torch.log(scores.clamp_min(self.eps))
        else:
            loss = 1.0 - scores

        # Dice loss is undefined for non-empty classes
        # So we zero contribution of channel that does not have true pixels
        # NOTE: A better workaround would be to use loss term `mean(y_pred)`
        # for this case, however it will be a modified jaccard loss

        mask = y_true.sum(dims) > 0
        loss *= mask.to(loss.dtype)

        if self.classes is not None:
            loss = loss[self.classes]

        return loss.mean()