模型

import torch
from torch import nn

__all__ = ['UNet', 'NestedUNet']


class VGGBlock(nn.Module):
    def __init__(self, in_channels, middle_channels, out_channels):
        super().__init__()
        self.relu = nn.ReLU(inplace=True)
        self.conv1 = nn.Conv2d(in_channels, middle_channels, 3, padding=1)
        self.bn1 = nn.BatchNorm2d(middle_channels)
        self.conv2 = nn.Conv2d(middle_channels, out_channels, 3, padding=1)
        self.bn2 = nn.BatchNorm2d(out_channels)

    def forward(self, x):
        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)
        out = self.relu(out)

        return out


class UNet(nn.Module):
    def __init__(self, num_classes, input_channels=3, **kwargs):
        super().__init__()

        nb_filter = [32, 64, 128, 256, 512]

        self.pool = nn.MaxPool2d(2, 2)
        self.up = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)#scale_factor:放大的倍数  插值

        self.conv0_0 = VGGBlock(input_channels, nb_filter[0], nb_filter[0])
        self.conv1_0 = VGGBlock(nb_filter[0], nb_filter[1], nb_filter[1])
        self.conv2_0 = VGGBlock(nb_filter[1], nb_filter[2], nb_filter[2])
        self.conv3_0 = VGGBlock(nb_filter[2], nb_filter[3], nb_filter[3])
        self.conv4_0 = VGGBlock(nb_filter[3], nb_filter[4], nb_filter[4])

        self.conv3_1 = VGGBlock(nb_filter[3]+nb_filter[4], nb_filter[3], nb_filter[3])
        self.conv2_2 = VGGBlock(nb_filter[2]+nb_filter[3], nb_filter[2], nb_filter[2])
        self.conv1_3 = VGGBlock(nb_filter[1]+nb_filter[2], nb_filter[1], nb_filter[1])
        self.conv0_4 = VGGBlock(nb_filter[0]+nb_filter[1], nb_filter[0], nb_filter[0])

        self.final = nn.Conv2d(nb_filter[0], num_classes, kernel_size=1)


    def forward(self, input):
        x0_0 = self.conv0_0(input)
        x1_0 = self.conv1_0(self.pool(x0_0))
        x2_0 = self.conv2_0(self.pool(x1_0))
        x3_0 = self.conv3_0(self.pool(x2_0))
        x4_0 = self.conv4_0(self.pool(x3_0))

        x3_1 = self.conv3_1(torch.cat([x3_0, self.up(x4_0)], 1))
        x2_2 = self.conv2_2(torch.cat([x2_0, self.up(x3_1)], 1))
        x1_3 = self.conv1_3(torch.cat([x1_0, self.up(x2_2)], 1))
        x0_4 = self.conv0_4(torch.cat([x0_0, self.up(x1_3)], 1))

        output = self.final(x0_4)
        return output


class NestedUNet(nn.Module):
    def __init__(self, num_classes, input_channels=3, deep_supervision=False, **kwargs):
        super().__init__()

        nb_filter = [32, 64, 128, 256, 512]

        self.deep_supervision = deep_supervision

        self.pool = nn.MaxPool2d(2, 2)
        self.up = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)

        self.conv0_0 = VGGBlock(input_channels, nb_filter[0], nb_filter[0])
        self.conv1_0 = VGGBlock(nb_filter[0], nb_filter[1], nb_filter[1])
        self.conv2_0 = VGGBlock(nb_filter[1], nb_filter[2], nb_filter[2])
        self.conv3_0 = VGGBlock(nb_filter[2], nb_filter[3], nb_filter[3])
        self.conv4_0 = VGGBlock(nb_filter[3], nb_filter[4], nb_filter[4])

        self.conv0_1 = VGGBlock(nb_filter[0]+nb_filter[1], nb_filter[0], nb_filter[0])
        self.conv1_1 = VGGBlock(nb_filter[1]+nb_filter[2], nb_filter[1], nb_filter[1])
        self.conv2_1 = VGGBlock(nb_filter[2]+nb_filter[3], nb_filter[2], nb_filter[2])
        self.conv3_1 = VGGBlock(nb_filter[3]+nb_filter[4], nb_filter[3], nb_filter[3])

        self.conv0_2 = VGGBlock(nb_filter[0]*2+nb_filter[1], nb_filter[0], nb_filter[0])
        self.conv1_2 = VGGBlock(nb_filter[1]*2+nb_filter[2], nb_filter[1], nb_filter[1])
        self.conv2_2 = VGGBlock(nb_filter[2]*2+nb_filter[3], nb_filter[2], nb_filter[2])

        self.conv0_3 = VGGBlock(nb_filter[0]*3+nb_filter[1], nb_filter[0], nb_filter[0])
        self.conv1_3 = VGGBlock(nb_filter[1]*3+nb_filter[2], nb_filter[1], nb_filter[1])

        self.conv0_4 = VGGBlock(nb_filter[0]*4+nb_filter[1], nb_filter[0], nb_filter[0])

        if self.deep_supervision:
            self.final1 = nn.Conv2d(nb_filter[0], num_classes, kernel_size=1)
            self.final2 = nn.Conv2d(nb_filter[0], num_classes, kernel_size=1)
            self.final3 = nn.Conv2d(nb_filter[0], num_classes, kernel_size=1)
            self.final4 = nn.Conv2d(nb_filter[0], num_classes, kernel_size=1)
        else:
            self.final = nn.Conv2d(nb_filter[0], num_classes, kernel_size=1)


    def forward(self, input):
        # print('input:',input.shape)
        x0_0 = self.conv0_0(input)
        # print('x0_0:',x0_0.shape)
        x1_0 = self.conv1_0(self.pool(x0_0))
        # print('x1_0:',x1_0.shape)
        x0_1 = self.conv0_1(torch.cat([x0_0, self.up(x1_0)], 1))
        # print('x0_1:',x0_1.shape)

        x2_0 = self.conv2_0(self.pool(x1_0))
        # print('x2_0:',x2_0.shape)
        x1_1 = self.conv1_1(torch.cat([x1_0, self.up(x2_0)], 1))
        # print('x1_1:',x1_1.shape)
        x0_2 = self.conv0_2(torch.cat([x0_0, x0_1, self.up(x1_1)], 1))
        # print('x0_2:',x0_2.shape)

        x3_0 = self.conv3_0(self.pool(x2_0))
        # print('x3_0:',x3_0.shape)
        x2_1 = self.conv2_1(torch.cat([x2_0, self.up(x3_0)], 1))
        # print('x2_1:',x2_1.shape)
        x1_2 = self.conv1_2(torch.cat([x1_0, x1_1, self.up(x2_1)], 1))
        # print('x1_2:',x1_2.shape)
        x0_3 = self.conv0_3(torch.cat([x0_0, x0_1, x0_2, self.up(x1_2)], 1))
        # print('x0_3:',x0_3.shape)
        x4_0 = self.conv4_0(self.pool(x3_0))
        # print('x4_0:',x4_0.shape)
        x3_1 = self.conv3_1(torch.cat([x3_0, self.up(x4_0)], 1))
        # print('x3_1:',x3_1.shape)
        x2_2 = self.conv2_2(torch.cat([x2_0, x2_1, self.up(x3_1)], 1))
        # print('x2_2:',x2_2.shape)
        x1_3 = self.conv1_3(torch.cat([x1_0, x1_1, x1_2, self.up(x2_2)], 1))
        # print('x1_3:',x1_3.shape)
        x0_4 = self.conv0_4(torch.cat([x0_0, x0_1, x0_2, x0_3, self.up(x1_3)], 1))
        # print('x0_4:',x0_4.shape)

        if self.deep_supervision:
            output1 = self.final1(x0_1)
            output2 = self.final2(x0_2)
            output3 = self.final3(x0_3)
            output4 = self.final4(x0_4)
            return [output1, output2, output3, output4]

        else:
            output = self.final(x0_4)
            return output

损失函数

BCEDiceLoss：

• 这个损失函数结合了二元交叉熵损失（Binary Cross Entropy, BCE）和 Dice Loss。
• BCE 于衡量模型输出和真实标签之间的二值化像素级别匹配情况。
• Dice Loss 用于量模型输出和真实标签之间的相似度，但这里采用了一种稍微不同的计算方式，即将 Dice Loss 作为 1 减去 Dice 相似度的平均值，这样得到的损失越小，说明相似度越高。

LovaszHingeLoss：

• 这个损失函数采用的是 Lovasz-Hinge Loss，它是一种用于处理不平衡数据集的损失函数，尤其适用于像素级别的分类任务。
• Lovasz-Hinge Loss 能够更好地处理类别不平衡和边界情况，相比于交叉熵损失，在处理不平衡数据时更加稳定。
  LovaszHingeLoss相关介绍

测试用例：

lovasz_losses.py 相关内容

"""
Lovasz-Softmax and Jaccard hinge loss in PyTorch
Maxim Berman 2018 ESAT-PSI KU Leuven (MIT License)
"""

from __future__ import print_function, division

import torch
from torch.autograd import Variable
import torch.nn.functional as F
import numpy as np

try:
    from itertools import ifilterfalse
except ImportError:  # py3k
    from itertools import filterfalse as ifilterfalse


def lovasz_grad(gt_sorted):
    """
    Computes gradient of the Lovasz extension w.r.t sorted errors
    See Alg. 1 in paper
    """
    p = len(gt_sorted)
    gts = gt_sorted.sum()
    intersection = gts - gt_sorted.float().cumsum(0)
    union = gts + (1 - gt_sorted).float().cumsum(0)
    jaccard = 1. - intersection / union
    if p > 1:  # cover 1-pixel case
        jaccard[1:p] = jaccard[1:p] - jaccard[0:-1]
    return jaccard


def iou_binary(preds, labels, EMPTY=1., ignore=None, per_image=True):
    """
    IoU for foreground class
    binary: 1 foreground, 0 background
    """
    if not per_image:
        preds, labels = (preds,), (labels,)
    ious = []
    for pred, label in zip(preds, labels):
        intersection = ((label == 1) & (pred == 1)).sum()
        union = ((label == 1) | ((pred == 1) & (label != ignore))).sum()
        if not union:
            iou = EMPTY
        else:
            iou = float(intersection) / float(union)
        ious.append(iou)
    iou = mean(ious)  # mean accross images if per_image
    return 100 * iou


def iou(preds, labels, C, EMPTY=1., ignore=None, per_image=False):
    """
    Array of IoU for each (non ignored) class
    """
    if not per_image:
        preds, labels = (preds,), (labels,)
    ious = []
    for pred, label in zip(preds, labels):
        iou = []
        for i in range(C):
            if i != ignore:  # The ignored label is sometimes among predicted classes (ENet - CityScapes)
                intersection = ((label == i) & (pred == i)).sum()
                union = ((label == i) | ((pred == i) & (label != ignore))).sum()
                if not union:
                    iou.append(EMPTY)
                else:
                    iou.append(float(intersection) / float(union))
        ious.append(iou)
    ious = [mean(iou) for iou in zip(*ious)]  # mean accross images if per_image
    return 100 * np.array(ious)


# --------------------------- BINARY LOSSES ---------------------------


def lovasz_hinge(logits, labels, per_image=True, ignore=None):
    """
    Binary Lovasz hinge loss
      logits: [B, H, W] Variable, logits at each pixel (between -\infty and +\infty)
      labels: [B, H, W] Tensor, binary ground truth masks (0 or 1)
      per_image: compute the loss per image instead of per batch
      ignore: void class id
    """
    if per_image:
        loss = mean(lovasz_hinge_flat(*flatten_binary_scores(log.unsqueeze(0), lab.unsqueeze(0), ignore))
                    for log, lab in zip(logits, labels))
    else:
        loss = lovasz_hinge_flat(*flatten_binary_scores(logits, labels, ignore))
    return loss


def lovasz_hinge_flat(logits, labels):
    """
    Binary Lovasz hinge loss
      logits: [P] Variable, logits at each prediction (between -\infty and +\infty)
      labels: [P] Tensor, binary ground truth labels (0 or 1)
      ignore: label to ignore
    """
    if len(labels) == 0:
        # only void pixels, the gradients should be 0
        return logits.sum() * 0.
    signs = 2. * labels.float() - 1.
    errors = (1. - logits * Variable(signs))
    errors_sorted, perm = torch.sort(errors, dim=0, descending=True)
    perm = perm.data
    gt_sorted = labels[perm]
    grad = lovasz_grad(gt_sorted)
    loss = torch.dot(F.relu(errors_sorted), Variable(grad))
    return loss


def flatten_binary_scores(scores, labels, ignore=None):
    """
    Flattens predictions in the batch (binary case)
    Remove labels equal to 'ignore'
    """
    scores = scores.view(-1)
    labels = labels.view(-1)
    if ignore is None:
        return scores, labels
    valid = (labels != ignore)
    vscores = scores[valid]
    vlabels = labels[valid]
    return vscores, vlabels


class StableBCELoss(torch.nn.modules.Module):
    def __init__(self):
        super(StableBCELoss, self).__init__()

    def forward(self, input, target):
        neg_abs = - input.abs()
        loss = input.clamp(min=0) - input * target + (1 + neg_abs.exp()).log()
        return loss.mean()


def binary_xloss(logits, labels, ignore=None):
    """
    Binary Cross entropy loss
      logits: [B, H, W] Variable, logits at each pixel (between -\infty and +\infty)
      labels: [B, H, W] Tensor, binary ground truth masks (0 or 1)
      ignore: void class id
    """
    logits, labels = flatten_binary_scores(logits, labels, ignore)
    loss = StableBCELoss()(logits, Variable(labels.float()))
    return loss


# --------------------------- MULTICLASS LOSSES ---------------------------


def lovasz_softmax(probas, labels, classes='present', per_image=False, ignore=None):
    """
    Multi-class Lovasz-Softmax loss
      probas: [B, C, H, W] Variable, class probabilities at each prediction (between 0 and 1).
              Interpreted as binary (sigmoid) output with outputs of size [B, H, W].
      labels: [B, H, W] Tensor, ground truth labels (between 0 and C - 1)
      classes: 'all' for all, 'present' for classes present in labels, or a list of classes to average.
      per_image: compute the loss per image instead of per batch
      ignore: void class labels
    """
    if per_image:
        loss = mean(lovasz_softmax_flat(*flatten_probas(prob.unsqueeze(0), lab.unsqueeze(0), ignore), classes=classes)
                    for prob, lab in zip(probas, labels))
    else:
        loss = lovasz_softmax_flat(*flatten_probas(probas, labels, ignore), classes=classes)
    return loss


def lovasz_softmax_flat(probas, labels, classes='present'):
    """
    Multi-class Lovasz-Softmax loss
      probas: [P, C] Variable, class probabilities at each prediction (between 0 and 1)
      labels: [P] Tensor, ground truth labels (between 0 and C - 1)
      classes: 'all' for all, 'present' for classes present in labels, or a list of classes to average.
    """
    if probas.numel() == 0:
        # only void pixels, the gradients should be 0
        return probas * 0.
    C = probas.size(1)
    losses = []
    class_to_sum = list(range(C)) if classes in ['all', 'present'] else classes
    for c in class_to_sum:
        fg = (labels == c).float()  # foreground for class c
        if (classes == 'present' and fg.sum() == 0):
            continue
        if C == 1:
            if len(classes) > 1:
                raise ValueError('Sigmoid output possible only with 1 class')
            class_pred = probas[:, 0]
        else:
            class_pred = probas[:, c]
        errors = (Variable(fg) - class_pred).abs()
        errors_sorted, perm = torch.sort(errors, 0, descending=True)
        perm = perm.data
        fg_sorted = fg[perm]
        losses.append(torch.dot(errors_sorted, Variable(lovasz_grad(fg_sorted))))
    return mean(losses)


def flatten_probas(probas, labels, ignore=None):
    """
    Flattens predictions in the batch
    """
    if probas.dim() == 3:
        # assumes output of a sigmoid layer
        B, H, W = probas.size()
        probas = probas.view(B, 1, H, W)
    B, C, H, W = probas.size()
    probas = probas.permute(0, 2, 3, 1).contiguous().view(-1, C)  # B * H * W, C = P, C
    labels = labels.view(-1)
    if ignore is None:
        return probas, labels
    valid = (labels != ignore)
    vprobas = probas[valid.nonzero().squeeze()]
    vlabels = labels[valid]
    return vprobas, vlabels


def xloss(logits, labels, ignore=None):
    """
    Cross entropy loss
    """
    return F.cross_entropy(logits, Variable(labels), ignore_index=255)


# --------------------------- HELPER FUNCTIONS ---------------------------
def isnan(x):
    return x != x


def mean(l, ignore_nan=False, empty=0):
    """
    nanmean compatible with generators.
    """
    l = iter(l)
    if ignore_nan:
        l = ifilterfalse(isnan, l)
    try:
        n = 1
        acc = next(l)
    except StopIteration:
        if empty == 'raise':
            raise ValueError('Empty mean')
        return empty
    for n, v in enumerate(l, 2):
        acc += v
    if n == 1:
        return acc
    return acc / n

import torch
import torch.nn as nn
import torch.nn.functional as F
from lovasz_losses import lovasz_hinge

# __all__ = ['BCEDiceLoss', 'LovaszHingeLoss']


class BCEDiceLoss(nn.Module):
    def __init__(self):
        super().__init__()

    def forward(self, input, target):


        bce = F.binary_cross_entropy_with_logits(input, target)
        smooth = 1e-5
        input = torch.sigmoid(input)
        num = target.size(0)
        input = input.view(num, -1)
        target = target.view(num, -1)
        intersection = (input * target)
        dice = (2. * intersection.sum(1) + smooth) / (input.sum(1) + target.sum(1) + smooth)
        dice = 1 - dice.sum() / num
        return 0.5 * bce + dice


class LovaszHingeLoss(nn.Module):
    def __init__(self):
        super().__init__()

    def forward(self, input, target):
        input = input.squeeze(1)
        target = target.squeeze(1)
        loss = lovasz_hinge(input, target, per_image=True)

        return loss


if __name__ == '__main__':
    import torch

    # 假设模型输出和真实标签都是二值化的图像，大小为(1, H, W)
    output = torch.tensor([[[0.3, 0.7], [0.8, 0.6]]])  # 模型输出
    # output = output.round().long()


    target = torch.tensor([[[0, 1], [1, 0]]],dtype=torch.float)  # 真实标签

    bce_dice_loss = BCEDiceLoss()
    bce_dice = bce_dice_loss(output, target)

    lovasz_hinge_loss = LovaszHingeLoss()
    lovasz_hinge = lovasz_hinge_loss(output, target)

    print("BCE Dice Loss:", bce_dice)
    print("Lovasz Hinge Loss:", lovasz_hinge)

原理解释和数学公式：

BCEDiceLoss 原理：

• BCE Dice Loss 结合了二元交叉熵损失和 Dice Loss。其数学表达式如下：

$$BCE\_Dice\_Loss=0.5\times BCE+(1-Dice)$$

其中，$BCE$ 表示二元交叉熵损失，$Dice$ 表示 Dice 相似度。这个损失函数的目标是最小化二元交叉熵损失和最大化 Dice 相似度，以达到更好的模型训练效果。

LovaszHingeLoss 原理：

• Lovasz-Hinge Loss 是一种非平衡数据集上的损失函数，用于像素级别的分类任务。其数学表达式如下：

$$Lovasz\_Hinge\_Loss=\text{lovasz\_hinge}(input,target)$$

这里的 $\text{lovasz\_hinge}$ 是一个函数，用于计算 Lovasz-Hinge Loss。

训练

评估函数

metrics.py

import numpy as np
import torch
import torch.nn.functional as F


def iou_score(output, target):
    smooth = 1e-5

    if torch.is_tensor(output):
        output = torch.sigmoid(output).data.cpu().numpy()
    if torch.is_tensor(target):
        target = target.data.cpu().numpy()
    output_ = output > 0.5
    target_ = target > 0.5
    intersection = (output_ & target_).sum()
    union = (output_ | target_).sum()

    return (intersection + smooth) / (union + smooth)


def dice_coef(output, target):
    smooth = 1e-5

    output = torch.sigmoid(output).view(-1).data.cpu().numpy()
    target = target.view(-1).data.cpu().numpy()
    intersection = (output * target).sum()

    return (2. * intersection + smooth) / \
        (output.sum() + target.sum() + smooth)


if __name__ == '__main__':
    import numpy as np
    import torch

    # 假设模型输出和真实标签都是二值化的图像，大小为(1, H, W)
    output = torch.tensor([[[0.3, 0.7], [0.8, 0.6]]])  # 模型输出
    target = torch.tensor([[[0, 1], [1, 0]]])  # 真实标签

    iou = iou_score(output, target)
    dice = dice_coef(output, target)

    print("IoU Score:", iou)
    print("Dice Coefficient:", dice)

IoU（Intersection over Union）评分函数原理

IoU 是一种常用的图像分割评价指标，它衡量了模型输出与真实标签之间的重程度。其数学公式如下：

$$IoU=\frac{TP}{TP+FP+FN}$$

其中，$TP$ 表示真正例（模型正确预测为正样本的数量），$FP$ 表示假正例（模型错误预测为正样本的数量），$FN$ 表示假负例（模型错误预测为负样本的数量）。

Dice Coefficient评分函数原理

Dice Coefficient 也是一种常用的图像分割评价指标，衡量模型输出和真实标签之间的相似度。其数学公式如下：

$$Dice=\frac{2\times TP}{2\times TP+FP+FN}$$

其中，$TP$ 表示真正例，$FP$ 表示假正例，$FN$ 表示假负例，与 IoU 公式中的定义相同。

这两个评分函数都以模型的真正例为分子，而分母则是真正例、假正例和假负例的总和，以此来衡量模型预测结果与真实标签的相似程度。公式中的平滑因子用于避免分母为零的情况，增加了数值稳定性。