loss.py

import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from torch.autograd import Variable
from collections import OrderedDict
from custom_transforms import edge_contour


def one_hot(index_tensor, cls_num):
    b, h, w = index_tensor.size()
    index_tensor = index_tensor.view(b, 1, h, w)
    one_hot_tensor = torch.cuda.FloatTensor(b, cls_num, h, w).zero_()
    one_hot_tensor = one_hot_tensor.cuda(index_tensor.get_device())
    target = one_hot_tensor.scatter_(1, index_tensor.long(), 1)

    return target


class NLLMultiLabelSmooth(nn.Module):
    def __init__(self, smoothing=0.1, nclasses=6):
        super(NLLMultiLabelSmooth, self).__init__()
        self.confidence = 1.0 - smoothing
        self.smoothing = smoothing
        self.nclasses = nclasses
        self.log_softmax = nn.LogSoftmax(dim=1)

    def forward(self, x, target):
        if self.training:
            x = x.float()
            target = target.float()
            target = one_hot(target, self.nclasses)
            logprobs = self.log_softmax(x)
            nll_loss = -logprobs * target

            nll_loss = nll_loss.sum(1)
            smooth_loss = -logprobs.mean(1)
            loss = self.confidence * nll_loss + self.smoothing * smooth_loss

            return loss.mean()
        else:
            return nn.CrossEntropyLoss(x, target)


class Edge_weak_loss(nn.Module):
    def __init__(self):
        super(Edge_weak_loss, self).__init__()
        self.ce_loss = nn.CrossEntropyLoss(ignore_index=255)

    def forward(self, scale_pred, target):
        edge = edge_contour(target).long()
        edge_loss = (torch.mul(self.ce_loss(scale_pred, target), torch.where(
            edge == 0, torch.tensor([1.]).cuda(), torch.tensor([2.0]).cuda()))).mean()

        return edge_loss


class Edge_loss(nn.Module):

    def __init__(self, ignore_index=255):
        super(Edge_loss, self).__init__()
        self.ignore_index = ignore_index

    def forward(self, pred, label):
        # h, w = label.size(1), label.size(2)
        pos_num = torch.sum(label == 1, dtype=torch.float)
        neg_num = torch.sum(label == 0, dtype=torch.float)
        weight_pos = neg_num / (pos_num + neg_num)
        weight_neg = pos_num / (pos_num + neg_num)

        weights = torch.Tensor([weight_neg, weight_pos])
        edge_loss = F.cross_entropy(pred, label,
                                weights.cuda(), ignore_index=self.ignore_index)

        return edge_loss


class EdgeLabelSmoothing(nn.Module):
    '''
    Cite: https://github.com/lin-honghui/tianchi_CountyAgriculturalBrain_top1/ @ lin.honghui@qq.com
    Description: 借鉴标签平滑的思想，针对样本中的预假设的 `hard sample`（图像边缘、不同类别交界） 进行标签平滑；
                 平滑因子可指定 smoothing 固定，或在训练过程中，在图像边缘、类间交界设置一定大小过渡带，统计过渡带
                 内像素 `hard sample` 比例动态调整。
    Args (type):
        win_size (int): 过渡带窗口大小；
        num_classes (int): 总类别数目，本次实验类别数为5；
        smoothing (float): 默认值为0.1，若指定 fix_smoothing ，则固定训练过程固定平滑因子为 smoothing。
    '''

    def __init__(self, win_size=11, num_classes=4, smoothing=0.1, fix_smoothing=False):
        super(EdgeLabelSmoothing, self).__init__()
        self.fix_smoothing = fix_smoothing
        assert (win_size % 2) == 1
        self.smoothing = smoothing / (num_classes - 1)
        self.win_size = win_size
        self.num_classes = num_classes

        self.find_edge_Conv = nn.Conv2d(in_channels=1, out_channels=1, kernel_size=win_size,
                                        padding=(win_size - 1) // 2, stride=1, bias=False)
        self.find_edge_Conv.weight.requires_grad = False
        new_state_dict = OrderedDict()
        weight = torch.zeros(1, 1, win_size, win_size)
        weight = weight - 1
        weight[:, :, win_size // 2, win_size // 2] = win_size * win_size - 1
        new_state_dict['weight'] = weight
        self.find_edge_Conv.load_state_dict(new_state_dict)

    def to_categorical(self, y, alpha=0.05, num_classes=None):
        y = np.array(y, dtype='int')
        input_shape = y.shape
        if input_shape and input_shape[-1] == 1 and len(input_shape) > 1:
            input_shape = tuple(input_shape[:-1])
        y = y.ravel()
        if not num_classes:
            num_classes = np.max(y) + 1
        n = y.shape[0]
        categorical = np.zeros((n, num_classes))
        categorical = categorical + alpha
        categorical[np.arange(n), y] = (1 - alpha) + (alpha / self.num_classes)
        output_shape = input_shape + (num_classes,)
        categorical = np.reshape(categorical, output_shape)
        categorical = categorical.transpose(0, 3, 1, 2)
        return categorical

    def forward(self, x, target):
        assert x.size(1) == self.num_classes
        log_p = nn.functional.log_softmax(x, dim=1)

        if not self.fix_smoothing:
            self.find_edge_Conv.cuda(device=target.device)
            edge_mask = self.find_edge_Conv(target)
            edge_mask = edge_mask.data.cpu().numpy()
            edge_mask[edge_mask != 0] = 1
            self.smoothing = np.mean(edge_mask)
            if self.smoothing > 0.2:
                self.smoothing = 0.2

        target = target.squeeze(dim=1)
        target = target.data.cpu().numpy()
        onehot_mask = self.to_categorical(target, 0, num_classes=self.num_classes)
        onehot_mask = onehot_mask * (1 - edge_mask)
        softlabel_mask = self.to_categorical(target, alpha=self.smoothing, num_classes=self.num_classes)
        softlabel_mask = softlabel_mask * edge_mask
        onehot_mask = torch.from_numpy(onehot_mask).cuda(device=log_p.device).float()
        softlabel_mask = torch.from_numpy(softlabel_mask).cuda(device=log_p.device).float()
        loss = torch.sum(onehot_mask * log_p + softlabel_mask * log_p, dim=1).mean()
        return -loss


class CrossEntropyLoss(nn.Module):
    def __init__(self, weights=None, ignore_index=255):
        super(CrossEntropyLoss, self).__init__()
        if weights is not None:
            weights = torch.from_numpy(np.array(weights)).float().cuda()
        self.ce_loss = nn.CrossEntropyLoss(ignore_index=ignore_index, weight=weights)

    def forward(self, prediction, label):
        loss = self.ce_loss(prediction, label)

        return loss


class CrossEntropyLoss_binary(nn.Module):
    def __init__(self, weights=None, binary_class=None, ignore_index=255):
        super(CrossEntropyLoss_binary, self).__init__()
        self.binary_class = binary_class
        if weights is not None:
            weights = torch.from_numpy(np.array(weights)).float().cuda()
        self.ce_loss = nn.CrossEntropyLoss(ignore_index=ignore_index, weight=weights)

    def forward(self, prediction, label):
        if self.binary_class is not None:
            label[label != self.binary_class] = 0
            label[label == self.binary_class] = 1
        loss = self.ce_loss(prediction, label)

        return loss


class DiceLoss(nn.Module):
    def __init__(self):
        super(DiceLoss, self).__init__()

    def forward(self, input, target):
        N = target.size(0)
        smooth = 1

        input_flat = input.view(N, -1)
        target_flat = target.view(N, -1)

        intersection = input_flat * target_flat

        loss = 2 * (intersection.sum(1) + smooth) / (input_flat.sum(1) + target_flat.sum(1) + smooth)
        loss = 1 - loss.sum() / N

        return loss


class MulticlassDiceLoss(nn.Module):
    """
    input: N C H W
    target: N H W

    """

    def __init__(self, weights=None, cls_num=None):
        super(MulticlassDiceLoss, self).__init__()
        self.weights = weights
        self.cls_num = cls_num

    def forward(self, input, target):

        target = one_hot(target, cls_num=self.cls_num)
        C = target.shape[1]

        # if weights is None:
        # 	weights = torch.ones(C) #uniform weights for all classes

        dice = DiceLoss()
        totalLoss = 0

        for i in range(C):
            diceLoss = dice(input[:, i], target[:, i])
            if self.weights is not None:
                diceLoss *= self.weights[i]
            totalLoss += diceLoss

        return totalLoss


class FocalLoss(nn.Module):
    def __init__(self, gamma=2, alpha=None, size_average=True):
        super(FocalLoss, self).__init__()
        self.gamma = gamma
        self.alpha = alpha
        if isinstance(alpha,(float,int)): self.alpha = torch.Tensor([alpha,1-alpha])
        if isinstance(alpha,list): self.alpha = torch.Tensor(alpha)
        self.size_average = size_average

    def forward(self, input, target):
        if input.dim()>2:
            input = input.view(input.size(0),input.size(1),-1)  # N,C,H,W => N,C,H*W
            input = input.transpose(1,2)    # N,C,H*W => N,H*W,C
            input = input.contiguous().view(-1,input.size(2))   # N,H*W,C => N*H*W,C
        target = target.view(-1,1)

        logpt = F.log_softmax(input)
        logpt = logpt.gather(1,target)
        logpt = logpt.view(-1)
        pt = Variable(logpt.data.exp())

        if self.alpha is not None:
            if self.alpha.type()!=input.data.type():
                self.alpha = self.alpha.type_as(input.data)
            at = self.alpha.gather(0,target.data.view(-1))
            logpt = logpt * Variable(at)

        loss = -1 * (1-pt)**self.gamma * logpt
        if self.size_average: return loss.mean()
        else: return loss.sum()


class Lovasz_loss(nn.Module):
    def __init__(self):
        super(Lovasz_loss, self).__init__()

    def lovasz_grad(self, 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 lovasz_softmax_flat(self, 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 is '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(self.lovasz_grad(fg_sorted))))
        return self.mean(losses)

    def flatten_probas(self, 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 isnan(self, x):
        return x != x

    def mean(self, l, ignore_nan=False, empty=0):
        """
        nanmean compatible with generators.
        """
        from itertools import filterfalse as ifilterfalse
        l = iter(l)
        if ignore_nan:
            l = ifilterfalse(self.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

    def forward(self, 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 = self.mean(
                self.lovasz_softmax_flat(*self.flatten_probas(prob.unsqueeze(0), lab.unsqueeze(0), ignore), classes=classes)
                for prob, lab in zip(probas, labels))
        else:
            loss = self.lovasz_softmax_flat(*self.flatten_probas(probas, labels, ignore), classes=classes)
        return loss