\u672c\u6587\u91cd\u70b9\u4e0d\u662f\u7b80\u5355\u7f57\u5217\u516c\u5f0f，\u800c\u662f\u56de\u7b54\u4e09\u4e2a\u95ee\u9898：<\/p>\n
\u8fd9\u4e2a Loss \u5230\u5e95\u5728\u4f18\u5316\u4ec0\u4e48？<\/span><\/p>\n<\/li>\n \u4e3a\u4ec0\u4e48\u8981\u8fd9\u6837\u8bbe\u8ba1？<\/span><\/p>\n<\/li>\n \u4ec0\u4e48\u65f6\u5019\u5e94\u8be5\u4f7f\u7528\u5b83？<\/span><\/p>\n<\/li>\n \u5728\u673a\u5668\u5b66\u4e60\u4e2d，\u6211\u4eec\u5e0c\u671b\u6a21\u578b\u00a0\u9884\u6d4b\u7ed3\u679c\u00a0 \u56e0\u6b64\u5b9a\u4e49\u4e00\u4e2a\u51fd\u6570：<\/p>\n \u7528\u6765\u8861\u91cf：\u6a21\u578b\u9884\u6d4b\u5f97\u6709\u591a\u5dee\u3002\u4e5f\u5c31\u662f\u00a0Loss\u00a0\u8d8a\u5c0f \u21d2 \u9884\u6d4b\u7ed3\u679c\u8d8a\u63a5\u8fd1\u76ee\u6807\u3002<\/p>\n \u8bad\u7ec3\u8fc7\u7a0b\u5c31\u662f\u4e0d\u65ad\u8c03\u6574\u6a21\u578b\u53c2\u6570 \u5176\u4e2d：<\/p>\n \u8fd9\u662f\u5b66\u4e60\u635f\u5931\u51fd\u6570\u65f6\u975e\u5e38\u5bb9\u6613\u6df7\u6dc6\u7684\u5730\u65b9\u3002<\/p>\n Loss \u7528\u4e8e：\u8bad\u7ec3\u6a21\u578b\u3002\u4f8b\u5982 pytorch \u4e2d\u7684\u5b9e\u73b0：<\/p>\n loss = nn.CrossEntropyLoss()<\/p>\n \u7136\u540e\u53cd\u5411\u4f20\u64ad\u66f4\u65b0\u6a21\u578b\u53c2\u6570：<\/p>\n loss.backward() Metric \u7528\u4e8e：\u8bc4\u4ef7\u6a21\u578b\u6548\u679c\u3002\u4f8b\u5982\u5206\u7c7b\u8bc4\u4ef7\u6307\u6807：<\/p>\n Accuracy<\/p>\n<\/li>\n Precision<\/p>\n<\/li>\n Recall<\/p>\n<\/li>\n F1<\/p>\n<\/li>\n AUC<\/p>\n<\/li>\n<\/ul>\n \u533b\u5b66\u56fe\u50cf\u5206\u5272\u8bc4\u4ef7\u6307\u6807：<\/p>\n Dice<\/p>\n<\/li>\n IoU<\/p>\n<\/li>\n HD95<\/p>\n<\/li>\n<\/ul>\n \u5177\u4f53\u516c\u5f0f\u4f8b\u5982：<\/p>\n Objective \u53ef\u4ee5\u7406\u89e3\u4e3a\u6700\u7ec8\u4f18\u5316\u76ee\u6807\u3002<\/p>\n \u4f8b\u5982\u591a\u4efb\u52a1\u5b66\u4e60：<\/p>\n \u8fd9\u91cc\u6574\u4e2a \u53ef\u4ee5\u628a\u6df1\u5ea6\u5b66\u4e60 Loss \u5927\u81f4\u5206\u6210\u4e0b\u9762\u51e0\u5927\u7c7b：<\/p>\n \u4e0b\u9762\u9010\u4e2a\u5c55\u5f00\u3002<\/p>\n \u56de\u5f52\u4efb\u52a1\u7684\u76ee\u6807\u901a\u5e38\u662f：<\/p>\n \u6700\u7ecf\u5178\u7684\u51e0\u79cd Loss \u662f：<\/p>\n MSE<\/p>\n<\/li>\n MAE<\/p>\n<\/li>\n Smooth L1<\/p>\n<\/li>\n Huber<\/p>\n<\/li>\n Log-Cosh<\/p>\n<\/li>\n<\/ul>\n MSE（Mean Squared Error，\u5747\u65b9\u8bef\u5dee）<\/p>\n \u5b9a\u4e49：<\/p>\n \u4e5f\u53ef\u4ee5\u5199\u6210：<\/p>\n \u5047\u8bbe：<\/p>\n \u6a21\u578b\u9884\u6d4b：<\/p>\n \u90a3\u4e48MSE：<\/p>\n \u5982\u679c\u9884\u6d4b\u9519\u5f97\u66f4\u591a：<\/p>\n \u90a3\u4e48：<\/p>\n \u6240\u4ee5 MSE \u4f1a：\u5f3a\u70c8\u60e9\u7f5a\u5927\u8bef\u5dee\u3002<\/p>\n \u5bf9\u4e8e：<\/p>\n \u5bf9 \u8bef\u5dee\u8d8a\u5927，\u68af\u5ea6\u8d8a\u5927\u3002<\/p>\n \u56e0\u6b64：MSE \u5bf9\u5f02\u5e38\u503c\u975e\u5e38\u654f\u611f\u3002<\/p>\n \u7b80\u5355<\/p>\n<\/li>\n \u5149\u6ed1<\/p>\n<\/li>\n \u53ef\u5bfc<\/p>\n<\/li>\n \u4f18\u5316\u7a33\u5b9a<\/p>\n<\/li>\n \u5927\u8bef\u5dee\u60e9\u7f5a\u5f3a<\/p>\n<\/li>\n<\/ul>\n PyTorch \u5b9e\u73b0：<\/p>\n import torch.nn as nn<\/p>\n criterion = nn.MSELoss()<\/p>\n loss = criterion(pred, target)<\/p>\n MAE：Mean Absolute Error<\/p>\n \u5b9a\u4e49：<\/p>\n \u4f8b\u5982：<\/p>\n \u5219：<\/p>\n \u5047\u8bbe：<\/p>\n MSE：<\/p>\n MAE：<\/p>\n \u56e0\u6b64：<\/p>\n MSE \u4f1a\u653e\u5927\u5927\u8bef\u5dee，\u800c MAE \u5bf9\u5f02\u5e38\u503c\u66f4\u52a0\u9c81\u68d2\u3002<\/p>\n \u7edd\u5bf9\u503c：<\/p>\n \u5728：<\/p>\n \u5904\u4e0d\u53ef\u5bfc<\/span>\u3002\u867d\u7136\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u80fd\u591f\u5904\u7406\u8fd9\u4e2a\u95ee\u9898，\u4f46\u76f8\u6bd4 MSE，\u5b83\u7684\u4f18\u5316\u901a\u5e38\u4e0d\u5982 MSE \u5e73\u6ed1\u3002<\/p>\n PyTorch \u5b9e\u73b0：<\/p>\n criterion = nn.L1Loss()<\/p>\n \u7b80\u5355\u7406\u89e3：<\/p>\n MSE：\u6211\u975e\u5e38\u8ba8\u538c\u5927\u9519\u8bef\u3002<\/p>\n MAE：\u5927\u9519\u8bef\u548c\u5c0f\u9519\u8bef\u6309\u7167\u7ebf\u6027\u6bd4\u4f8b\u5904\u7406\u3002<\/p>\n Smooth L1 \u662f\u4e00\u79cd\u00a0MSE \u548c MAE \u7684\u6298\u4e2d\u3002<\/p>\n \u5b9a\u4e49：<\/p>\n \u5176\u4e2d：<\/p>\n \u6838\u5fc3\u601d\u60f3：<\/p>\n \u5c0f\u8bef\u5dee：\u4f7f\u7528\u5e73\u65b9\u51fd\u6570<\/p>\n<\/li>\n \u5927\u8bef\u5dee：\u4f7f\u7528\u7edd\u5bf9\u503c\u51fd\u6570<\/p>\n<\/li>\n<\/ul>\n \u56e0\u6b64：\u5c0f\u8bef\u5dee\u533a\u57df\u5e73\u6ed1，\u5927\u8bef\u5dee\u533a\u57df\u9c81\u68d2\u3002<\/p>\n Huber Loss \u4e0e Smooth L1 \u975e\u5e38\u63a5\u8fd1\u3002<\/p>\n \u5b9a\u4e49：<\/p>\n \u5176\u4e2d：<\/p>\n \u5f53\u8bef\u5dee\u8f83\u5c0f\u65f6：<\/p>\n \u5f53\u8bef\u5dee\u8f83\u5927\u65f6：<\/p>\n \u56e0\u6b64：Huber Loss \u662f\u7ecf\u5178\u7684\u9c81\u68d2\u56de\u5f52 Loss\u3002<\/p>\n PyTorch \u5b9e\u73b0：<\/p>\n criterion = nn.HuberLoss(delta=1.0)<\/p>\n PyTorch \u4e5f\u5c06 SmoothL1Loss \u548c HuberLoss \u4f5c\u4e3a\u6807\u51c6 Loss \u63d0\u4f9b\u3002<\/p>\n \u5b9a\u4e49：<\/p>\n \u5f53 \u6240\u4ee5\u7c7b\u4f3c MSE\u3002<\/p>\n \u5f53\u00a0 \u6240\u4ee5\u7c7b\u4f3c MAE\u3002<\/p>\n \u56e0\u6b64：Log-Cosh \u4e5f\u662f\u4e00\u79cd\u5e73\u6ed1\u7684\u9c81\u68d2\u56de\u5f52 Loss\u3002<\/p>\n \u5bf9\u4e8e\u4e8c\u5206\u7c7b：<\/p>\n \u6a21\u578b\u8f93\u51fa：<\/p>\n BCE：<\/p>\n \u5982\u679c：<\/p>\n \u90a3\u4e48：<\/p>\n \u6240\u4ee5：<\/p>\n \u5f97\u5230：<\/p>\n \u5982\u679c\u6a21\u578b\u9519\u8bef\u5730\u9884\u6d4b：<\/p>\n \u90a3\u4e48：<\/p>\n Loss \u4f1a\u975e\u5e38\u5927\u3002<\/p>\n \u5b9e\u9645\u6df1\u5ea6\u5b66\u4e60\u4e2d\u66f4\u63a8\u8350：<\/p>\n nn.BCEWithLogitsLoss()<\/p>\n \u800c\u4e0d\u662f：<\/p>\n sigmoid() \u56e0\u4e3a BCEWithLogitsLoss \u5c06 Sigmoid \u4e0e BCE \u5408\u5e76\u8ba1\u7b97，\u5e76\u4f7f\u7528\u6570\u503c\u7a33\u5b9a\u6280\u5de7，\u907f\u514d\u6982\u7387\u975e\u5e38\u63a5\u8fd1 0 \u6216 1 \u65f6\u51fa\u73b0\u6570\u503c\u95ee\u9898\u3002<\/p>\n \u6b63\u786e\u5199\u6cd5：<\/p>\n criterion = nn.BCEWithLogitsLoss()<\/p>\n logits = model(x)<\/p>\n loss = criterion(logits, target)<\/p>\n \u6a21\u578b\u6700\u540e：<\/p>\n nn.Linear(…)<\/p>\n \u4e0d\u8981\u518d\u624b\u52a8 Sigmoid\u3002<\/p>\n BCE \u975e\u5e38\u9002\u5408：\u4e8c\u5206\u7c7b<\/span><\/p>\n \u4f8b\u5982：<\/p>\n \u6b63\u5e38 \/ \u60a3\u75c5 \u4f8b\u5982\u4e00\u4e2a\u6837\u672c\u540c\u65f6\u5b58\u5728：<\/p>\n \u80ba\u708e \u6bcf\u4e2a\u6807\u7b7e\u90fd\u662f\u72ec\u7acb\u7684：<\/p>\n \u8fd9\u79cd\u60c5\u51b5\u4f7f\u7528：<\/p>\n nn.BCEWithLogitsLoss()<\/p>\n \u800c\u4e0d\u662f CrossEntropyLoss\u3002<\/p>\n Cross Entropy \u662f\u6df1\u5ea6\u5b66\u4e60\u5206\u7c7b\u4efb\u52a1\u4e2d\u6700\u91cd\u8981\u7684 Loss \u4e4b\u4e00\u3002<\/p>\n \u5bf9\u4e8e\u00a0 \u5982\u679c\u6807\u7b7e\u662f one-hot：<\/p>\n \u90a3\u4e48：<\/p>\n Softmax：<\/p>\n \u5176\u4e2d\u00a0 Cross Entropy：<\/p>\n \u4ee3\u5165 Softmax：<\/p>\n \u5f97\u5230：<\/p>\n \u8fd9\u5c31\u662f\u5206\u7c7b\u6a21\u578b\u5e38\u89c1\u7684 LogSumExp \u7ed3\u6784\u3002<\/p>\n PyTorch \u7684 CrossEntropyLoss <\/span>\u63a5\u6536\u7684\u662f \u672a\u5f52\u4e00\u5316 logits，\u800c\u4e0d\u662f\u5148 Softmax \u540e\u7684\u6982\u7387，\u5b83\u5185\u90e8\u7b49\u4ef7\u4e8e LogSoftmax + NLLLoss<\/span>\u3002<\/p>\n \u6240\u4ee5：<\/p>\n criterion = nn.CrossEntropyLoss()<\/p>\n logits = model(x)<\/p>\n loss = criterion(logits, target)<\/p>\n \u4e0d\u8981：<\/p>\n logits = model(x)<\/p>\n prob = torch.softmax(logits, dim=1)<\/p>\n loss = criterion(prob, target)<\/p>\n NLL：Negative Log Likelihood<\/p>\n \u5b9a\u4e49：<\/p>\n PyTorch：<\/p>\n criterion = nn.NLLLoss()<\/p>\n log_prob = F.log_softmax(logits, dim=1)<\/p>\n loss = criterion(log_prob, target)<\/p>\n \u56e0\u6b64：<\/p>\n \u5047\u8bbe\u771f\u5b9e\u7c7b\u522b：<\/p>\n \u6a21\u578b\u9884\u6d4b：<\/p>\n class 0: 0.1 Loss：<\/p>\n \u5982\u679c\u6a21\u578b\u9884\u6d4b：<\/p>\n class 0: 0.4 Loss：<\/p>\n \u6240\u4ee5：Cross Entropy \u672c\u8d28\u4e0a\u662f\u5728\u60e9\u7f5a\u6a21\u578b\u7ed9\u771f\u5b9e\u7c7b\u522b\u5206\u914d\u8fc7\u4f4e\u6982\u7387\u3002<\/p>\n \u666e\u901a One-Hot：<\/p>\n Label Smoothing \u540e：<\/p>\n \u4f8b\u5982：<\/p>\n \u53d8\u6210：<\/p>\n \u76ee\u7684：\u4e0d\u8ba9\u6a21\u578b\u8fc7\u5ea6\u81ea\u4fe1\u3002<\/p>\n PyTorch \u5b9e\u73b0：<\/p>\n criterion = nn.CrossEntropyLoss( \u8fd9\u662f\u975e\u5e38\u91cd\u8981\u7684\u77e5\u8bc6\u70b9\u3002<\/p>\n \u4f8b\u5982：<\/p>\n \u4e00\u5f20\u56fe\u7247\u53ea\u80fd\u662f：<\/p>\n \u732b \u4f7f\u7528：<\/p>\n CrossEntropyLoss<\/p>\n \u4e00\u5f20\u533b\u5b66\u5f71\u50cf\u53ef\u4ee5\u540c\u65f6：<\/p>\n \u80ba\u708e = 1 \u4f7f\u7528：<\/p>\n BCEWithLogitsLoss<\/p>\n KL：Kullback-Leibler Divergence<\/p>\n \u5b9a\u4e49：<\/p>\n \u5b83\u8861\u91cf：<\/p>\n \u4e24\u4e2a\u6982\u7387\u5206\u5e03\u4e4b\u95f4\u6709\u591a\u4e0d\u4e00\u6837\u3002<\/p>\n \u5047\u8bbe\u6559\u5e08\u6a21\u578b：<\/p>\n \u5b66\u751f\u6a21\u578b：<\/p>\n KL \u7528\u4e8e\u7ea6\u675f：<\/p>\n \u56e0\u6b64\u7ecf\u5e38\u7528\u4e8e：<\/p>\n Knowledge Distillation<\/p>\n<\/li>\n VAE<\/p>\n<\/li>\n \u5206\u5e03\u5bf9\u9f50<\/p>\n<\/li>\n \u6982\u7387\u6a21\u578b<\/p>\n<\/li>\n<\/ul>\n \u4ea4\u53c9\u71b5：<\/p>\n KL：<\/p>\n \u5c55\u5f00：<\/p>\n \u56e0\u4e3a Hinge Loss \u662f SVM \u7684\u7ecf\u5178\u635f\u5931\u3002<\/p>\n \u6807\u7b7e：<\/p>\n \u5b9a\u4e49：<\/p>\n \u5982\u679c：<\/p>\n \u90a3\u4e48：<\/p>\n \u5426\u5219：<\/p>\n \u5b83\u4e0d\u4ec5\u8981\u6c42\u5206\u7c7b\u6b63\u786e，\u8fd8\u8981\u6c42：\u9884\u6d4b\u7ed3\u679c\u62e5\u6709\u8db3\u591f\u5927\u7684 margin\u3002Hinge Loss \u662f\u6700\u5927\u95f4\u9694\u5206\u7c7b\u5668\u4e2d\u7684\u7ecf\u5178\u5f62\u5f0f\u3002<\/p>\n \u5728\u76ee\u6807\u68c0\u6d4b\u3001\u533b\u5b66\u5f71\u50cf\u7b49\u4efb\u52a1\u4e2d\u7ecf\u5e38\u5b58\u5728：\u6b63\u8d1f\u6837\u672c\u6781\u5ea6\u4e0d\u5e73\u8861<\/p>\n \u4f8b\u5982：<\/p>\n \u80cc\u666f：99.9% \u666e\u901a BCE \u5f88\u5bb9\u6613\u88ab\u5927\u91cf\u7b80\u5355\u8d1f\u6837\u672c\u652f\u914d\u3002<\/p>\n Focal Loss \u7684\u601d\u60f3\u662f：<\/p>\n \u964d\u4f4e\u7b80\u5355\u6837\u672c\u7684\u6743\u91cd，\u628a\u6ce8\u610f\u529b\u96c6\u4e2d\u5230\u56f0\u96be\u6837\u672c\u3002<\/p>\n \u5b9a\u4e49：<\/p>\n \u5176\u4e2d：<\/p>\n \u5047\u8bbe：<\/p>\n \u8fd9\u662f\u4e00\u4e2a\u975e\u5e38\u5bb9\u6613\u7684\u6837\u672c\u3002<\/p>\n \u5982\u679c：<\/p>\n \u5219：<\/p>\n \u8fd9\u4e2a\u6837\u672c\u7684\u8d21\u732e\u88ab\u6781\u5927\u964d\u4f4e\u3002<\/p>\n \u5982\u679c：<\/p>\n \u5219：<\/p>\n \u56f0\u96be\u6837\u672c\u4ecd\u7136\u5177\u6709\u8f83\u5927\u7684 Loss\u3002<\/p>\n \u6240\u4ee5：<\/p>\n FocalLoss = CrossEntropy + \u56f0\u96be\u6837\u672c\u52a0\u6743<\/p>\n import torch def focal_loss(logits, targets, alpha=0.25, gamma=2.0): probs = torch.sigmoid(logits)<\/p>\n pt = torch.where( alpha_t = torch.where( loss = alpha_t * (1 – pt).pow(gamma) * bce<\/p>\n return loss.mean()<\/p>\n Focal Loss \u5173\u6ce8：<\/p>\n \u6837\u672c\u96be\u5ea6\u3002<\/p>\n Class-Balanced Loss \u5173\u6ce8：<\/p>\n \u7c7b\u522b\u6570\u91cf\u4e0d\u5e73\u8861\u3002<\/p>\n \u4e00\u4e2a\u7ecf\u5178\u601d\u8def\u662f\u6839\u636e\u7c7b\u522b\u7684\u6709\u6548\u6837\u672c\u6570\u91cf：<\/p>\n \u5b9a\u4e49\u7c7b\u522b\u6743\u91cd：<\/p>\n \u5176\u4e2d \u6837\u672c\u8d8a\u5c11：<\/p>\n \u56e0\u6b64\u5c11\u6570\u7c7b\u4f1a\u5f97\u5230\u66f4\u5927\u7684\u6743\u91cd\u3002<\/p>\n Dice Loss \u662f\u533b\u5b66\u56fe\u50cf\u5206\u5272\u4e2d\u6781\u5176\u91cd\u8981\u7684 Loss\u3002<\/p>\n Dice：<\/p>\n \u5176\u4e2d：<\/p>\n Dice Loss：<\/p>\n \u901a\u5e38\u6dfb\u52a0\u5e73\u6ed1\u9879：<\/p>\n \u533b\u5b66\u56fe\u50cf\u901a\u5e38\u5b58\u5728：\u524d\u666f\u533a\u57df\u8fdc\u5c0f\u4e8e\u80cc\u666f\u533a\u57df\u3002<\/p>\n \u4f8b\u5982：<\/p>\n \u80ba\u90e8：90% \u751a\u81f3：<\/p>\n \u80cc\u666f：99% Pixel-wise BCE \u5f88\u5bb9\u6613\u88ab\u80cc\u666f\u4e3b\u5bfc\u3002<\/p>\n Dice \u76f4\u63a5\u8861\u91cf：\u9884\u6d4b\u533a\u57df\u548c\u771f\u5b9e\u533a\u57df\u7684\u91cd\u53e0\u7a0b\u5ea6，\u56e0\u6b64\u5bf9\u7c7b\u522b\u4e0d\u5e73\u8861\u66f4\u52a0\u53cb\u597d\u3002<\/p>\n def dice_loss(pred, target, smooth=1e-6): intersection = (pred * target).sum()<\/p>\n dice = ( return 1 – dice<\/p>\n \u533b\u5b66\u56fe\u50cf\u5206\u5272\u975e\u5e38\u5e38\u89c1：<\/p>\n \u4f8b\u5982：<\/p>\n loss = bce_loss + dice_loss<\/p>\n \u4e3a\u4ec0\u4e48\u7ec4\u5408？<\/p>\n BCE：<\/p>\n \u5173\u6ce8\u6bcf\u4e00\u4e2a\u50cf\u7d20\u3002<\/p>\n Dice：<\/p>\n \u5173\u6ce8\u6574\u4f53\u533a\u57df\u91cd\u53e0\u3002<\/p>\n \u6240\u4ee5\u4e8c\u8005\u5177\u6709\u4e92\u8865\u6027\u3002<\/p>\n IoU\u00a0：\u4e5f\u53eb\u00a0Jaccard Index<\/p>\n IoU Loss：<\/p>\n \u5199\u6210\u9884\u6d4b\u6982\u7387\u5f62\u5f0f：<\/p>\n \u4e8c\u8005\u5173\u7cfb：<\/p>\n \u53cd\u8fc7\u6765：<\/p>\n \u4f8b\u5982：<\/p>\n \u5219：<\/p>\n \u56e0\u6b64：Dice \u548c IoU \u672c\u8d28\u4e0a\u90fd\u5728\u8861\u91cf\u533a\u57df\u91cd\u53e0，\u4f46\u6570\u503c\u5e76\u4e0d\u76f8\u540c\u3002<\/p>\n Dice \u5bf9 FP \u548c FN \u7684\u60e9\u7f5a\u76f8\u5bf9\u5bf9\u79f0\u3002\u4f46\u533b\u5b66\u5206\u5272\u4e2d\u6709\u65f6：\u6f0f\u8bca\u6bd4\u8bef\u68c0\u66f4\u52a0\u4e25\u91cd\u3002<\/p>\n \u4e8e\u662f\u53ef\u4ee5\u4f7f\u7528 Tversky Index：<\/p>\n \u5176\u4e2d：<\/p>\n Tversky Loss：<\/p>\n \u5982\u679c：<\/p>\n \u90a3\u4e48：<\/p>\n \u53d7\u5230\u66f4\u5927\u7684\u60e9\u7f5a\u3002<\/p>\n \u4e5f\u5c31\u662f：\u66f4\u5173\u6ce8 Recall，\u51cf\u5c11\u6f0f\u68c0\u3002<\/p>\n \u5982\u679c：<\/p>\n \u5219\u66f4\u52a0\u60e9\u7f5a：<\/p>\n \u5373：\u66f4\u5173\u6ce8 Precision，\u51cf\u5c11\u8bef\u68c0\u3002<\/p>\n Tversky \u8fd8\u53ef\u4ee5\u4e0e Focal \u601d\u60f3\u7ed3\u5408：<\/p>\n \u8fd9\u6837：\u5bf9\u56f0\u96be\u533a\u57df\u8fdb\u884c\u8fdb\u4e00\u6b65\u5f3a\u8c03\u3002\u533b\u5b66\u56fe\u50cf\u4e2d\u5c0f\u76ee\u6807\u75c5\u7076\u5206\u5272\u7ecf\u5e38\u53ef\u4ee5\u8003\u8651\u8fd9\u79cd\u8bbe\u8ba1\u3002<\/p>\n Dice\/IoU \u4e3b\u8981\u5173\u6ce8：\u533a\u57df\u91cd\u53e0\u3002\u4f46\u662f\u5bf9\u4e8e\u4e00\u4e9b\u533b\u5b66\u56fe\u50cf：\u8fb9\u754c\u975e\u5e38\u91cd\u8981\u3002<\/p>\n \u4f8b\u5982：<\/p>\n \u80bf\u7624\u8fb9\u754c \u6b64\u65f6\u53ef\u4ee5\u5f15\u5165 Boundary Loss\u3002\u5176\u6838\u5fc3\u601d\u60f3\u662f：\u76f4\u63a5\u4f18\u5316\u9884\u6d4b\u8fb9\u754c\u4e0e\u771f\u5b9e\u8fb9\u754c\u4e4b\u95f4\u7684\u8ddd\u79bb\u3002<\/p>\n \u4f8b\u5982\u53ef\u4ee5\u4f7f\u7528\u8ddd\u79bb\u53d8\u6362：<\/p>\n \u5176\u4e2d：<\/p>\n \u8868\u793a\u50cf\u7d20 \u533b\u5b66\u56fe\u50cf\u5206\u5272\u7ecf\u5e38\u5173\u6ce8：Hausdorff Distance\u3002<\/p>\n Hausdorff Distance：<\/p>\n \u5b83\u5173\u6ce8：\u6700\u574f\u60c5\u51b5\u4e0b\u7684\u8fb9\u754c\u8ddd\u79bb\u3002<\/p>\n \u56e0\u6b64：<\/p>\n Dice \u9ad8<\/p>\n<\/li>\n \u4f46\u67d0\u4e2a\u8fb9\u754c\u4f4d\u7f6e\u504f\u5dee\u5f88\u5927<\/p>\n<\/li>\n<\/ul>\n \u4ecd\u7136\u53ef\u80fd\u5f97\u5230\u8f83\u5927\u7684 HD，\u8fd9\u4e5f\u662f\u4e3a\u4ec0\u4e48\u533b\u5b66\u5206\u5272\u4e0d\u80fd\u53ea\u770b Dice\u3002<\/p>\n \u76ee\u6807\u68c0\u6d4b\u901a\u5e38\u540c\u65f6\u9700\u8981\u4f18\u5316：<\/p>\n \u5176\u4e2d：<\/p>\n \u5206\u7c7b Loss<\/p>\n<\/li>\n Objectness Loss<\/p>\n<\/li>\n Bounding Box Regression Loss<\/p>\n<\/li>\n<\/ul>\n Bounding Box Loss \u7684\u7ecf\u5178\u65b9\u6cd5\u5305\u62ec：<\/p>\n L1<\/p>\n<\/li>\n Smooth L1<\/p>\n<\/li>\n IoU Loss<\/p>\n<\/li>\n GIoU Loss<\/p>\n<\/li>\n DIoU Loss<\/p>\n<\/li>\n CIoU Loss<\/p>\n<\/li>\n<\/ul>\n \u4e24\u4e2a\u6846：<\/p>\n IoU：<\/p>\n Loss：<\/p>\n \u95ee\u9898\u662f：\u5f53\u4e24\u4e2a\u6846\u6ca1\u6709\u91cd\u53e0\u65f6，IoU = 0\u3002\u6b64\u65f6\u4e0d\u540c\u4f4d\u7f6e\u7684\u6846：<\/p>\n IoU = 0 \u65e0\u6cd5\u544a\u8bc9\u6a21\u578b：\u5230\u5e95\u5e94\u8be5\u5f80\u54ea\u4e2a\u65b9\u5411\u79fb\u52a8\u3002<\/p>\n GIoU：<\/p>\n \u5176\u4e2d： Loss：<\/p>\n \u56e0\u6b64\u5373\u4f7f：<\/p>\n GIoU \u4ecd\u7136\u80fd\u591f\u63d0\u4f9b\u4e00\u5b9a\u68af\u5ea6\u4fe1\u606f\u3002<\/p>\n DIoU \u4e0d\u4ec5\u8003\u8651\u91cd\u53e0\u9762\u79ef，\u8fd8\u8003\u8651：\u4e24\u4e2a\u6846\u4e2d\u5fc3\u70b9\u4e4b\u95f4\u7684\u8ddd\u79bb\u3002<\/p>\n \u5b9a\u4e49：<\/p>\n \u5176\u4e2d：<\/p>\n \u6240\u4ee5：DIoU \u4f1a\u76f4\u63a5\u63a8\u52a8\u9884\u6d4b\u6846\u4e2d\u5fc3\u5411\u771f\u5b9e\u6846\u4e2d\u5fc3\u79fb\u52a8\u3002<\/p>\n CIoU \u5728 DIoU \u57fa\u7840\u4e0a\u8fdb\u4e00\u6b65\u8003\u8651：<\/p>\n Overlap<\/p>\n<\/li>\n Center distance<\/p>\n<\/li>\n Aspect ratio<\/p>\n<\/li>\n<\/ul>\n \u5b9a\u4e49：<\/p>\n \u5176\u4e2d：<\/p>\n \u56e0\u6b64：CIoU \u540c\u65f6\u4f18\u5316\u4f4d\u7f6e\u3001\u91cd\u53e0\u548c\u5bbd\u9ad8\u6bd4\u3002<\/p>\n \u53ef\u4ee5\u8bb0\u6210：<\/p>\n IoU \u6240\u4ee5：<\/p>\n \u5bf9\u6bd4\u5b66\u4e60\u7684\u76ee\u6807\u4e0d\u662f\u76f4\u63a5\u9884\u6d4b\u7c7b\u522b，\u800c\u662f\u5b66\u4e60：\u4ec0\u4e48\u6837\u7684\u6837\u672c\u5e94\u8be5\u63a5\u8fd1，\u4ec0\u4e48\u6837\u7684\u6837\u672c\u5e94\u8be5\u8fdc\u79bb\u3002<\/p>\n \u7ed9\u5b9a\u4e24\u4e2a\u6837\u672c：<\/p>\n \u6807\u7b7e：<\/p>\n Contrastive Loss：<\/p>\n \u5176\u4e2d：<\/p>\n \u5982\u679c\u4e24\u4e2a\u6837\u672c\u662f\u6b63\u6837\u672c：<\/p>\n \u90a3\u4e48：<\/p>\n \u8bad\u7ec3\u4f1a\u8ba9：<\/p>\n \u5373：\u6b63\u6837\u672c\u8d8a\u6765\u8d8a\u63a5\u8fd1\u3002<\/p>\n \u5982\u679c\u662f\u8d1f\u6837\u672c：<\/p>\n \u5219：<\/p>\n \u53ea\u6709：<\/p>\n \u624d\u6709 Loss\u3002<\/p>\n \u56e0\u6b64：\u8d1f\u6837\u672c\u53ea\u9700\u8981\u8ddd\u79bb\u8d85\u8fc7 margin \u5373\u53ef\u3002<\/p>\n Triplet Loss \u4f7f\u7528\u4e09\u4e2a\u6837\u672c：<\/p>\n Anchor \u5206\u522b\u8bb0\u4e3a：<\/p>\n \u76ee\u6807：<\/p>\n \u5b9a\u4e49：<\/p>\n \u5176\u4e2d： PyTorch \u7684 TripletMarginLoss \u5c31\u91c7\u7528\u8fd9\u4e00\u57fa\u672c\u5f62\u5f0f，\u7528 Anchor\u3001Positive\u3001Negative \u4e09\u5143\u7ec4\u4f18\u5316\u76f8\u5bf9\u8ddd\u79bb\u3002<\/p>\n \u5047\u8bbe：<\/p>\n margin：<\/p>\n \u5219：<\/p>\n \u8bf4\u660e\u5df2\u7ecf\u6ee1\u8db3\u8981\u6c42\u3002<\/p>\n \u5982\u679c：<\/p>\n \u5219：<\/p>\n \u9700\u8981\u7ee7\u7eed\u4f18\u5316\u3002<\/p>\n \u5bf9\u4e8e\u4e24\u4e2a\u5411\u91cf：<\/p>\n Cosine Similarity：<\/p>\n Cosine Embedding Loss \u7684\u76ee\u6807\u662f：\u6b63\u6837\u672c cosine similarity \u9ad8，\u8d1f\u6837\u672c cosine similarity \u4f4e\u3002<\/p>\n \u975e\u5e38\u9002\u5408：<\/p>\n \u6587\u672c\u5339\u914d<\/p>\n<\/li>\n \u56fe\u6587\u5bf9\u9f50<\/p>\n<\/li>\n Sentence Embedding<\/p>\n<\/li>\n \u591a\u6a21\u6001\u5b66\u4e60<\/p>\n<\/li>\n<\/ul>\n Ranking Loss \u7528\u4e8e：<\/p>\n \u6392\u5e8f\u4efb\u52a1\u3002<\/p>\n \u4f8b\u5982：<\/p>\n \u5546\u54c1 A \u5e94\u8be5\u6392\u5728\u5546\u54c1 B \u524d\u9762<\/p>\n \u5b9a\u4e49：<\/p>\n \u5176\u4e2d：<\/p>\n \u4f8b\u5982：<\/p>\n \u8868\u793a：<\/p>\n CTC：Connectionist Temporal Classification<\/p>\n \u4e3b\u8981\u7528\u4e8e：<\/p>\n OCR<\/p>\n<\/li>\n ASR<\/p>\n<\/li>\n \u624b\u5199\u8bc6\u522b<\/p>\n<\/li>\n \u65e0\u660e\u786e\u5bf9\u9f50\u5173\u7cfb\u7684\u5e8f\u5217\u9884\u6d4b<\/p>\n<\/li>\n<\/ul>\n \u4f8b\u5982\u8f93\u5165：<\/p>\n \u56fe\u50cf\u7279\u5f81： \u76ee\u6807：<\/p>\n HELLO<\/p>\n \u4f46\u662f\u6211\u4eec\u5e76\u4e0d\u77e5\u9053：<\/p>\n H \u5bf9\u5e94 t1\/t2？ CTC \u901a\u8fc7：\u5bf9\u6240\u6709\u53ef\u80fd\u7684\u5bf9\u9f50\u8def\u5f84\u6c42\u548c\u00a0\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u3002<\/p>\n \u5b9a\u4e49：<\/p>\n \u5176\u4e2d：<\/p>\n Loss：<\/p>\n \u56e0\u6b64：CTC \u4e0d\u9700\u8981\u4eba\u5de5\u63d0\u4f9b\u6bcf\u4e2a\u65f6\u95f4\u6b65\u5bf9\u5e94\u54ea\u4e2a\u5b57\u7b26\u3002<\/p>\n VAE：Variational Autoencoder<\/p>\n VAE \u7684 Loss \u901a\u5e38\u7531\u4e24\u90e8\u5206\u7ec4\u6210：<\/p>\n \u5373：<\/p>\n VAE Loss \u4f8b\u5982\u56fe\u50cf\u91cd\u6784：<\/p>\n \u5b83\u8981\u6c42：<\/p>\n \u5373：\u7f16\u7801\u4e4b\u540e\u4ecd\u7136\u80fd\u591f\u6062\u590d\u539f\u59cb\u8f93\u5165\u3002<\/p>\n VAE \u5e0c\u671b：<\/p>\n \u63a5\u8fd1：<\/p>\n \u901a\u5e38：<\/p>\n \u56e0\u6b64：<\/p>\n \u5982\u679c：<\/p>\n \u5219：<\/p>\n \u6700\u7ec8：<\/p>\n GAN \u5305\u62ec：<\/p>\n \u7ecf\u5178 GAN：<\/p>\n \u7ecf\u5178 GAN \u4e2d：<\/p>\n \u5b9e\u9645\u8bad\u7ec3\u4e2d\u7ecf\u5e38\u4f7f\u7528 Non-Saturating Loss：<\/p>\n \u76ee\u6807：\u8ba9 Discriminator \u628a\u751f\u6210\u6837\u672c\u5224\u65ad\u6210\u771f\u3002<\/p>\n \u6269\u6563\u6a21\u578b\u7684\u6838\u5fc3\u662f：\u5411\u6570\u636e\u4e2d\u4e0d\u65ad\u52a0\u5165\u566a\u58f0，\u518d\u8bad\u7ec3\u6a21\u578b\u9884\u6d4b\u566a\u58f0\u3002<\/p>\n \u524d\u5411\u8fc7\u7a0b：<\/p>\n \u5176\u4e2d：<\/p>\n \u6a21\u578b：<\/p>\n \u9884\u6d4b\u566a\u58f0\u3002\u6700\u7ecf\u5178\u7684\u8bad\u7ec3 Loss：<\/p>\n \u4e5f\u5c31\u662f：\u8ba9\u6a21\u578b\u9884\u6d4b\u81ea\u5df1\u52a0\u8fdb\u53bb\u7684\u566a\u58f0\u3002<\/p>\n \u4e25\u683c\u6765\u8bf4，L1\/L2 \u4e0d\u4e00\u5b9a\u662f\u4efb\u52a1 Loss，\u800c\u662f：\u53c2\u6570\u6b63\u5219\u5316\u9879\u3002<\/p>\n \u4f5c\u7528：\u9f13\u52b1\u53c2\u6570\u53d8\u6210 0\u3002\u56e0\u6b64\u53ef\u4ee5\u4ea7\u751f：\u7a00\u758f\u6a21\u578b\u3002<\/p>\n \u6700\u7ec8：<\/p>\n \u4f5c\u7528：\u6291\u5236\u6a21\u578b\u53c2\u6570\u8fc7\u5927，\u964d\u4f4e\u8fc7\u62df\u5408\u98ce\u9669\u3002<\/p>\n Elastic Net \u540c\u65f6\u7ed3\u5408：<\/p>\n \u6240\u4ee5：<\/p>\n L1 \u2192 \u7a00\u758f \u73b0\u5b9e\u4efb\u52a1\u5f80\u5f80\u4e0d\u662f\u53ea\u6709\u4e00\u4e2a\u76ee\u6807\u3002<\/p>\n \u4f8b\u5982\u533b\u5b66 AI：<\/p>\n \u8f93\u5165 CT \u5206\u522b\u5f97\u5230：<\/p>\n \u6700\u7ec8：<\/p>\n \u5047\u8bbe：<\/p>\n \u4f46\u662f：<\/p>\n \u76f4\u63a5\u76f8\u52a0：<\/p>\n \u8fd9\u610f\u5473\u7740：Segmentation Loss \u5bf9\u68af\u5ea6\u7684\u5f71\u54cd\u53ef\u80fd\u8fdc\u5927\u4e8e Classification Loss\u3002<\/p>\n \u56e0\u6b64\u9700\u8981：<\/p>\n \u4f8b\u5982：<\/p>\n \u5b9e\u9645\u79d1\u7814\u4e2d，\u5355\u72ec\u4f7f\u7528\u4e00\u4e2a Loss \u5e76\u4e0d\u4e00\u5b9a\u662f\u6700\u597d\u7684\u9009\u62e9\u3002<\/p>\n \u5e38\u89c1\u7ec4\u5408\u5305\u62ec：<\/p>\n \u6216\u8005：<\/p>\n \u8fd9\u662f\u4e00\u4e2a\u975e\u5e38\u5b9e\u7528\u7684\u95ee\u9898\u3002\u5047\u8bbe\u75c5\u7076\u53ea\u53601%，\u80cc\u666f\u536099%<\/p>\n \u5982\u679c\u4f7f\u7528 BCE：<\/p>\n \u80cc\u666f\u50cf\u7d20\u975e\u5e38\u591a \u800c Dice：<\/p>\n \u76f4\u63a5\u5173\u6ce8\u533a\u57df\u91cd\u53e0 \u6240\u4ee5：<\/p>\n \u53ef\u4ee5\u540c\u65f6：<\/p>\n \u5229\u7528 CE \u7684\u50cf\u7d20\u7ea7\u76d1\u7763<\/p>\n<\/li>\n \u5229\u7528 Dice \u7684\u533a\u57df\u7ea7\u76d1\u7763<\/p>\n<\/li>\n<\/ul>\n \u53ef\u4ee5\u8bb0\u4f4f\u4e0b\u9762\u8fd9\u5f20\u8868\u3002<\/p>\n \u4e0d\u4e00\u5b9a\u3002<\/p>\n \u4f8b\u5982\u8bad\u7ec3：<\/p>\n \u4f46\u662f：<\/p>\n \u8bf4\u660e：<\/p>\n \u53ef\u80fd\u53d1\u751f\u8fc7\u62df\u5408\u3002<\/p>\n \u751a\u81f3：<\/p>\n \u4f46：<\/p>\n \u4e5f\u53ef\u80fd\u51fa\u73b0\u3002<\/p>\n \u56e0\u6b64\u4e00\u5b9a\u8981\u533a\u5206：<\/p>\n Training Loss \u56e0\u4e3a Loss \u548c\u6700\u7ec8\u4e1a\u52a1\u76ee\u6807\u53ef\u80fd\u4e0d\u540c\u3002\u4f8b\u5982\u533b\u5b66\u5206\u5272：<\/p>\n \u8bad\u7ec3：<\/p>\n \u4f46\u662f\u8bba\u6587\u6700\u7ec8\u62a5\u544a：<\/p>\n Dice \u56e0\u4e3a：<\/p>\n Loss \u662f\u4f18\u5316\u5de5\u5177，Metric \u662f\u8bc4\u4ef7\u5de5\u5177\u3002<\/p>\n \u4e8c\u8005\u6ca1\u6709\u5fc5\u8981\u5b8c\u5168\u4e00\u81f4\u3002<\/p>\n \u6df1\u5ea6\u5b66\u4e60\u771f\u6b63\u66f4\u65b0\u7684\u662f：<\/p>\n \u53c2\u6570\u66f4\u65b0：<\/p>\n \u56e0\u6b64\u8bbe\u8ba1 Loss \u65f6\u4e0d\u80fd\u53ea\u770b：<\/p>\n Loss \u6570\u503c\u662f\u5426\u5408\u7406\u3002<\/p>\n \u66f4\u91cd\u8981\u7684\u662f：<\/p>\n \u68af\u5ea6\u662f\u5426\u80fd\u591f\u63d0\u4f9b\u6709\u6548\u7684\u4f18\u5316\u65b9\u5411\u3002<\/p>\n \u4f8b\u5982 Sigmoid：<\/p>\n \u5176\u5bfc\u6570：<\/p>\n \u5f53：<\/p>\n \u6216\u8005：<\/p>\n \u90fd\u6709：<\/p>\n \u56e0\u6b64：<\/p>\n Sigmoid \u9971\u548c\u4f1a\u5bfc\u81f4\u68af\u5ea6\u53d8\u5c0f\u3002<\/p>\n \u8fd9\u4e5f\u662f\u4e3a\u4ec0\u4e48\u5b9e\u9645\u4e8c\u5206\u7c7b\u4e2d\u63a8\u8350：<\/p>\n BCEWithLogitsLoss<\/p>\n \u800c\u4e0d\u662f\u624b\u52a8：<\/p>\n Sigmoid + BCELoss<\/p>\n PyTorch Loss \u7ecf\u5e38\u6709：<\/p>\n reduction="mean"<\/p>\n \u6216\u8005：<\/p>\n reduction="sum"<\/p>\n \u6216\u8005：<\/p>\n reduction="none"<\/p>\n \u5047\u8bbe：<\/p>\n loss = [1, 2, 3, 4]<\/p>\n \u56e0\u4e3a\u5f88\u591a\u9ad8\u7ea7 Loss \u9700\u8981：<\/p>\n \u5148\u8ba1\u7b97\u6bcf\u4e2a\u6837\u672c\u7684 Loss，\u518d\u8fdb\u884c\u52a0\u6743\u3002<\/p>\n \u4f8b\u5982 Focal：<\/p>\n loss = criterion( \u7136\u540e：<\/p>\n loss = weight * loss<\/p>\n \u6700\u540e：<\/p>\n loss = loss.mean()<\/p>\n \u8fd9\u4e5f\u662f\u5b9e\u73b0\u81ea\u5b9a\u4e49 Loss \u7684\u5e38\u89c1\u65b9\u5f0f\u3002<\/p>\n PyTorch \u5f53\u524d torch.nn \u63d0\u4f9b\u4e86\u5927\u91cf\u6807\u51c6\u635f\u5931，\u5305\u62ec：<\/p>\n L1Loss BCELoss CrossEntropyLoss KLDivLoss<\/p>\n MarginRankingLoss CosineEmbeddingLoss<\/p>\n TripletMarginLoss CTCLoss<\/p>\n PoissonNLLLoss \u8fd9\u4e9b\u90fd\u53ef\u4ee5\u5728 PyTorch \u5b98\u65b9 API \u4e2d\u627e\u5230\u3002<\/p>\n \u53ef\u4ee5\u628a Loss \u770b\u6210\u4e0b\u9762\u8fd9\u68f5\u6811：<\/p>\n Loss Function \u4e0d\u8981\u6b7b\u8bb0\u516c\u5f0f\u3002<\/p>\n \u5efa\u8bae\u6309\u7167\u4e0b\u9762\u56db\u4e2a\u95ee\u9898\u7406\u89e3\u4efb\u4f55\u4e00\u4e2a Loss\u3002<\/p>\n \u4f8b\u5982：<\/p>\n MSE：<\/p>\n Dice：<\/p>\n KL：<\/p>\n Triplet：<\/p>\n MSE：<\/p>\n \u5927\u8bef\u5dee\u3002<\/p>\n Focal：<\/p>\n \u56f0\u96be\u6837\u672c\u3002<\/p>\n Dice：<\/p>\n \u533a\u57df\u4e0d\u91cd\u53e0\u3002<\/p>\n Tversky：<\/p>\n FP\/FN\u3002<\/p>\n CIoU：<\/p>\n \u6846\u7684\u4f4d\u7f6e\u3001\u91cd\u53e0\u548c\u5f62\u72b6\u3002<\/p>\n \u4f8b\u5982：<\/p>\n MSE Huber Focal Dice Tversky CIoU \u53ef\u4ee5\u628a\u6df1\u5ea6\u5b66\u4e60 Loss \u7684\u6f14\u5316\u7406\u89e3\u6210：<\/p>\n \u8bef\u5dee\u2192\u6982\u7387\u2192\u7c7b\u522b\u2192\u96be\u6837\u672c\u2192\u533a\u57df\u2192\u8fb9\u754c\u2192\u8ddd\u79bb\u2192\u5206\u5e03\u2192\u591a\u4efb\u52a1<\/p>\n \u6700\u57fa\u7840\u7684\u662f：<\/p>\n \u5206\u7c7b：<\/p>\n \u4e0d\u5e73\u8861：<\/p>\n \u5206\u5272：<\/p>\n \u68c0\u6d4b：<\/p>\n \u8868\u5f81\u5b66\u4e60：<\/p>\n \u5206\u5e03\u5b66\u4e60：<\/p>\n \u751f\u6210\u6a21\u578b：<\/p>\n \u6700\u7ec8\u590d\u6742\u6a21\u578b\u5f80\u5f80\u4e0d\u662f\u4f7f\u7528\u5355\u4e00 Loss，\u800c\u662f：<\/p>\n \u56e0\u6b64\u771f\u6b63\u7684 Loss \u8bbe\u8ba1，\u672c\u8d28\u4e0a\u662f\u5728\u56de\u7b54：<\/p>\n \u201c\u6211\u5e0c\u671b\u6a21\u578b\u5177\u6709\u4ec0\u4e48\u6837\u7684\u884c\u4e3a？\u201d<\/p>\n \u5982\u679c\u5e0c\u671b\u6a21\u578b\u9884\u6d4b\u6570\u503c\u51c6\u786e，\u5c31\u4f7f\u7528\u56de\u5f52 Loss；\u5982\u679c\u5e0c\u671b\u5206\u7c7b\u6982\u7387\u51c6\u786e，\u5c31\u4f7f\u7528 Cross Entropy；\u5982\u679c\u5e0c\u671b\u5173\u6ce8\u56f0\u96be\u6837\u672c，\u5c31\u4f7f\u7528 Focal；\u5982\u679c\u5e0c\u671b\u9884\u6d4b\u533a\u57df\u91cd\u53e0，\u5c31\u4f7f\u7528 Dice\/IoU；\u5982\u679c\u5e0c\u671b\u5b66\u4e60\u8868\u793a\u4e4b\u95f4\u7684\u76f8\u4f3c\u5173\u7cfb，\u5c31\u4f7f\u7528 Contrastive\/Triplet；\u5982\u679c\u540c\u65f6\u5b58\u5728\u591a\u4e2a\u76ee\u6807，\u5c31\u7ec4\u5408\u591a\u4e2a Loss\u3002<\/p>\n \u8fd9\u4e5f\u662f\u6df1\u5ea6\u5b66\u4e60\u4e2d\u4e00\u4e2a\u975e\u5e38\u91cd\u8981\u7684\u601d\u60f3：<\/p>\n Loss\u00a0Function \u2248 \u6211\u4eec\u5bf9\u201c\u4ec0\u4e48\u53eb\u505a\u9884\u6d4b\u5f97\u597d\u201d\u7684\u6570\u5b66\u5b9a\u4e49<\/p>\n \u53c2\u8003\u8d44\u6599<\/p>\n PyTorch torch.nn Loss \u5b98\u65b9\u6587\u6863：\u5305\u542b L1\u3001MSE\u3001BCE\u3001CrossEntropy\u3001KL\u3001CTC\u3001Triplet \u7b49\u6807\u51c6\u5b9e\u73b0\u3002<\/p>\n<\/li>\n PyTorch CrossEntropyLoss：\u8be6\u7ec6\u8bf4\u660e logits\u3001class index\u3001class probability\u3001weight\u3001ignore_index \u548c label_smoothing \u7b49\u673a\u5236\u3002<\/p>\n<\/li>\n PyTorch BCEWithLogitsLoss：Sigmoid \u4e0e BCE \u7684\u6570\u503c\u7a33\u5b9a\u5b9e\u73b0，\u5e76\u652f\u6301 pos_weight \u5904\u7406\u6b63\u8d1f\u6837\u672c\u4e0d\u5e73\u8861\u3002<\/p>\n<\/li>\n PyTorch TripletMarginLoss：Anchor\u3001Positive\u3001Negative \u4e09\u5143\u7ec4\u53ca margin-based \u8ddd\u79bb\u4f18\u5316\u3002<\/p>\n<\/li>\n scikit-learn Hinge Loss \u6587\u6863：\u4ecb\u7ecd\u4e8c\u5206\u7c7b\u548c\u591a\u5206\u7c7b Hinge Loss\u3002<\/p>\n<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":" \u672c\u6587\u7cfb\u7edf\u603b\u7ed3\u6df1\u5ea6\u5b66\u4e60\u4e2d\u5e38\u89c1\u7684\u635f\u5931\u51fd\u6570，\u4ece\u6700\u57fa\u7840\u7684 MSE\u3001MAE\u3001\u4ea4\u53c9\u71b5，\u5230\u5206\u7c7b\u4e2d\u7684 BCE\u3001Focal Loss，\u76ee\u6807\u68c0\u6d4b\u4e2d\u7684 IoU \/\u00a0GIoU \/ DIoU \/ CIoU，\u8bed\u4e49\u5206\u5272\u4e2d\u7684 Dice Loss\u3001Tversky Loss，\u518d\u5230\u5bf9\u6bd4\u5b66\u4e60\u4e2d\u7684 Contrastive Loss\u3001Triplet Loss，\u4ee5\u53ca KL Divergence\u3001CTC\u3001GAN\u3001\u6269\u6563\u6a21\u578b\u4e2d\u7684\u635f\u5931\u51fd\u6570\u3002\u672c\u6587\u91cd<\/p>\n","protected":false},"author":2,"featured_media":99964,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[50,86],"topic":[],"class_list":["post-99984","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-50","tag-86"],"yoast_head":"\n
\n1. \u4ec0\u4e48\u662f\u635f\u5931\u51fd\u6570？<\/h2>\n
![\"$\\\\hat{y}$\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260903090350-6a9937f68c845.png\")
\u00a0\u5c3d\u53ef\u80fd\u63a5\u8fd1\u00a0\u771f\u5b9e\u6807\u7b7e\u00a0![\"$y$\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260903090350-6a9937f6a3afa.png\")
\u3002<\/p>\n![\"\\\\mathcal{L}(y,\\\\hat{y})\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260903090350-6a9937f6a8d0d.png\")
<\/p>\n![\"$\\\\theta$\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260903090350-6a9937f6adf1f.png\")
：<\/p>\n![\"\\\\boxed{](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260903090350-6a9937f6b31a3.png\")
<\/p>\n\n
![\"$x$\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260903090350-6a9937f6b97ca.png\")
：\u8f93\u5165<\/p>\n<\/li>\n：\u771f\u5b9e\u6807\u7b7e<\/p>\n<\/li>\n![\"$f_\\\\theta(x)$\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260903090350-6a9937f6be9e9.png\")
：\u6a21\u578b\u9884\u6d4b，\u5373\u00a0![\"\\\\hat{y}\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260903090350-6a9937f6c3ba8.png\")
<\/p>\n<\/li>\n：\u6a21\u578b\u53c2\u6570<\/p>\n<\/li>\n![\"$\\\\mathcal{L}$\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260903090350-6a9937f6c8e9e.png\")
：\u635f\u5931\u51fd\u6570<\/p>\n<\/li>\n<\/ul>\n
\n2. Loss\u3001Metric \u548c Objective \u6709\u4ec0\u4e48\u533a\u522b？<\/h2>\n
2.1 Loss<\/h3>\n
\noptimizer.step()<\/p>\n
\n2.2 Metric<\/h3>\n
\n
\n
![\"Dice=](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260903090350-6a9937f6ce37b.png\")
<\/p>\n
\n2.3 Objective<\/h3>\n
![\"\\\\mathcal{L}](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260903090350-6a9937f6d486e.png\")
<\/p>\n\u5c31\u662f\u6700\u7ec8 Objective\u3002<\/p>\n
\n3. Loss \u51fd\u6570\u7684\u603b\u4f53\u5206\u7c7b<\/h2>\n
\n
\u7c7b\u522b\u5178\u578b Loss\u5178\u578b\u4efb\u52a1<\/tr>\n \n\n \u56de\u5f52 Loss<\/td>\n MSE\u3001MAE\u3001Huber<\/td>\n \u56de\u5f52<\/td>\n<\/tr>\n \n \u6982\u7387\u5206\u5e03 Loss<\/td>\n KL Divergence<\/td>\n \u84b8\u998f\u3001\u5206\u5e03\u5339\u914d<\/td>\n<\/tr>\n \n \u4e8c\u5206\u7c7b Loss<\/td>\n BCE\u3001BCEWithLogits<\/td>\n \u4e8c\u5206\u7c7b\u3001\u591a\u6807\u7b7e<\/td>\n<\/tr>\n \n \u591a\u5206\u7c7b Loss<\/td>\n CE\u3001NLL<\/td>\n \u591a\u5206\u7c7b<\/td>\n<\/tr>\n \n Margin Loss<\/td>\n Hinge\u3001Margin Ranking<\/td>\n SVM\u3001\u6392\u5e8f<\/td>\n<\/tr>\n \n \u4e0d\u5e73\u8861 Loss<\/td>\n Focal\u3001Class-Balanced<\/td>\n \u957f\u5c3e\u5206\u7c7b<\/td>\n<\/tr>\n \n \u5206\u5272 Loss<\/td>\n Dice\u3001IoU\u3001Tversky<\/td>\n \u533b\u5b66\u56fe\u50cf\u5206\u5272<\/td>\n<\/tr>\n \n \u68c0\u6d4b Loss<\/td>\n IoU\u3001GIoU\u3001DIoU\u3001CIoU<\/td>\n \u76ee\u6807\u68c0\u6d4b<\/td>\n<\/tr>\n \n \u5ea6\u91cf\u5b66\u4e60<\/td>\n Contrastive\u3001Triplet<\/td>\n \u8868\u5f81\u5b66\u4e60<\/td>\n<\/tr>\n \n \u5e8f\u5217 Loss<\/td>\n CTC<\/td>\n OCR\u3001\u8bed\u97f3<\/td>\n<\/tr>\n \n \u751f\u6210\u6a21\u578b<\/td>\n GAN Loss\u3001VAE Loss\u3001Diffusion Loss<\/td>\n \u751f\u6210\u6a21\u578b<\/td>\n<\/tr>\n \n \u6b63\u5219\u5316<\/td>\n L1\u3001L2\u3001Elastic Net<\/td>\n \u9632\u6b62\u8fc7\u62df\u5408<\/td>\n<\/tr>\n \n \u591a\u4efb\u52a1<\/td>\n Weighted Sum<\/td>\n \u591a\u4efb\u52a1\u5b66\u4e60<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
\n4. \u56de\u5f52\u635f\u5931\u51fd\u6570<\/h2>\n
![\"\\\\hat{y}\\\\approx](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260903090350-6a9937f6dc040.png\")
<\/p>\n\n
\n5. MSE Loss<\/h2>\n
![\"\\\\boxed{](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260903090350-6a9937f6e128a.png\")
<\/p>\n![\"L_{MSE}](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260903090350-6a9937f6e78b0.png\")
<\/p>\n
\n5.1 \u4e3a\u4ec0\u4e48\u53eb\u5e73\u65b9\u8bef\u5dee？<\/h3>\n
![\"y=10\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260903090350-6a9937f6edcdf.png\")
<\/p>\n![\"\\\\hat{y}=8\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260903090350-6a9937f6f2d47.png\")
<\/p>\n![\"L=(y-\\\\hat{y})^2=4\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260903090351-6a9937f703b59.png\")
<\/p>\n![\"\\\\hat{y}=5\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260903090351-6a9937f708b3e.png\")
<\/p>\n![\"L=(10-5)^2=25\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260903090351-6a9937f70db52.png\")
<\/p>\n
\n5.2 MSE \u7684\u68af\u5ea6<\/h3>\n
![\"L=(y-\\\\hat{y})^2\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260903090351-6a9937f713d93.png\")
<\/p>\n\u6c42\u5bfc：<\/p>\n![\"\\\\frac{\\\\partial](\"2026-09-034dcftgbsyun.png\")
<\/p>\n
\n5.3 MSE \u7684\u4f18\u70b9<\/h3>\n
\n
\n6. MAE \/ L1 Loss<\/h2>\n
![\"\\\\boxed{](\"2026-09-03mpxggiuaal5.png\")
<\/p>\n![\"y=10,\\\\quad](\"2026-09-03g0mim4u5ksc.png\")
<\/p>\n![\"L=|10-8|=2\"](\"2026-09-03bzu2bqhkldc.png\")
<\/p>\n
\n6.1 MAE \u548c MSE \u7684\u533a\u522b<\/h3>\n
![\"e=10\"](\"2026-09-03cqdsdhy3nj0.png\")
<\/p>\n![\"L=10^2=100\"](\"2026-09-03cjrzt5k5c12.png\")
<\/p>\n![\"L=|10|=10\"](\"2026-09-03legbz31nd23.png\")
<\/p>\n
\n6.2 MAE \u7684\u95ee\u9898<\/h3>\n
![\"|x|\"](\"2026-09-03ln4fzyl5ia5.png\")
<\/p>\n![\"x=0\"](\"2026-09-03jgvahn33qwm.png\")
<\/p>\n
\n7. MSE vs MAE<\/h2>\n
\n
\u7279\u6027MSEMAE<\/tr>\n \n\n \u8bef\u5dee\u5f62\u5f0f<\/td>\n ![\"$e^2$\"](\"2026-09-035yu2yjujlb1.png\")
<\/td>\n![\"|e|\"](\"2026-09-03aqafl4o4wzg.png\")
<\/td>\n<\/tr>\n\n \u5927\u8bef\u5dee\u60e9\u7f5a<\/td>\n \u5f3a<\/td>\n \u5f31<\/td>\n<\/tr>\n \n \u5f02\u5e38\u503c\u654f\u611f\u5ea6<\/td>\n \u9ad8<\/td>\n \u4f4e<\/td>\n<\/tr>\n \n \u68af\u5ea6<\/td>\n \u8fde\u7eed<\/td>\n 0 \u70b9\u4e0d\u53ef\u5bfc<\/td>\n<\/tr>\n \n \u4f18\u5316<\/td>\n \u5bb9\u6613<\/td>\n \u76f8\u5bf9\u56f0\u96be<\/td>\n<\/tr>\n \n \u9c81\u68d2\u6027<\/td>\n \u8f83\u5dee<\/td>\n \u8f83\u597d<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
\n8. Smooth L1 Loss<\/h2>\n
![\"L(x)=](\"2026-09-03mhjufylovsv.png\")
<\/p>\n![\"x=y-\\\\hat{y}\"](\"2026-09-03n5umxvszvlx.png\")
<\/p>\n\n
\n9. Huber Loss<\/h2>\n
![\"L_\\\\delta(r)](\"2026-09-03vi5vqtpiypm.png\")
<\/p>\n![\"r=y-\\\\hat{y}\"](\"2026-09-03b5ksilcjrcs.png\")
<\/p>\n![\"L\\\\sim](\"2026-09-031d2ecjdi1vd.png\")
<\/p>\n![\"L\\\\sim](\"2026-09-03nalotwwhzjx.png\")
<\/p>\n
\n10. Log-Cosh Loss<\/h2>\n
![\"\\\\boxed{](\"2026-09-03mwod3ta2laq.png\")
<\/p>\n \u5f88\u5c0f\u65f6：<\/p>\n![\"\\\\log\\\\cosh(x)\\\\approx\\\\frac{x^2}{2}\"](\"2026-09-03exiw12utkn3.png\")
<\/p>\n \u5f88\u5927\u65f6：<\/p>\n![\"\\\\log\\\\cosh(x)\\\\approx](\"2026-09-03kxm2x0msact.png\")
<\/p>\n
\n11. \u56de\u5f52 Loss \u603b\u7ed3<\/h2>\n
\n
Loss\u6838\u5fc3\u601d\u60f3\u5f02\u5e38\u503c<\/tr>\n \n\n MSE<\/td>\n \u5e73\u65b9\u8bef\u5dee<\/td>\n \u654f\u611f<\/td>\n<\/tr>\n \n MAE<\/td>\n \u7edd\u5bf9\u8bef\u5dee<\/td>\n \u9c81\u68d2<\/td>\n<\/tr>\n \n Smooth L1<\/td>\n MSE + MAE<\/td>\n \u9c81\u68d2<\/td>\n<\/tr>\n \n Huber<\/td>\n MSE + MAE<\/td>\n \u9c81\u68d2<\/td>\n<\/tr>\n \n Log-Cosh<\/td>\n \u5e73\u6ed1 MAE<\/td>\n \u8f83\u9c81\u68d2<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
\n12. \u4e8c\u5206\u7c7b\u635f\u5931：Binary Cross Entropy<\/h2>\n
![\"y\\\\in\\\\{0,1\\\\}\"](\"2026-09-03qw5f4m4unny.png\")
<\/p>\n![\"p=P(y=1|x)\"](\"2026-09-03c0aleoy2izp.png\")
<\/p>\n![\"\\\\boxed{](\"2026-09-03by0ahmjm0cx.png\")
<\/p>\n
\n\u4e3a\u4ec0\u4e48\u53ef\u4ee5\u5f53\u4f5c\u4e8c\u5206\u7c7b\u7684\u635f\u5931？<\/h3>\n
12.1 \u5f53\u771f\u5b9e\u6807\u7b7e\u662f 1<\/h4>\n
![\"y=1\"](\"2026-09-03cnebl5zal04.png\")
<\/p>\n![\"L=-\\\\log](\"2026-09-031emwfwa0zll.png\")
<\/p>\n\n
![\"$p=0.99$\"](\"2026-09-030vxgqdy5pnl.png\")
\u00a0\u2192 Loss \u5f88\u5c0f<\/p>\n<\/li>\n![\"$p=0.9$\"](\"2026-09-03xe4yxchv5g4.png\")
\u00a0 \u2192 Loss \u8f83\u5c0f<\/p>\n<\/li>\n![\"$p=0.1$\"](\"2026-09-032ajy0km4b2n.png\")
\u00a0\u2192 Loss \u5f88\u5927<\/p>\n<\/li>\n<\/ul>\n
\n12.2 \u5f53\u771f\u5b9e\u6807\u7b7e\u662f 0<\/h4>\n
![\"y=0\"](\"2026-09-03wcaikw2h3xd.png\")
<\/p>\n![\"L=-\\\\log(1-p)\"](\"2026-09-032dpgvxcjivi.png\")
<\/p>\n![\"p=0.99\"](\"2026-09-03wjv11zubm5d.png\")
<\/p>\n![\"L=-\\\\log(0.01)\"](\"2026-09-03k1efu2kgiyf.png\")
<\/p>\n
\n13. BCEWithLogitsLoss<\/h2>\n
\n+
\nBCELoss()<\/p>\n
\n14. BCE \u6700\u9002\u5408\u4ec0\u4e48\u4efb\u52a1？<\/h2>\n
\n\u732b \/ \u72d7
\n\u6709\u76ee\u6807 \/ \u65e0\u76ee\u6807<\/p>\n\u591a\u6807\u7b7e\u5206\u7c7b<\/h4>\n
\n\u80f8\u8154\u79ef\u6db2
\n\u80ba\u4e0d\u5f20
\n\u6c14\u80f8<\/p>\n![\"y_i\\\\in\\\\{0,1\\\\}\"](\"2026-09-033ylhus0uz2v.png\")
<\/p>\n
\n15. Cross Entropy Loss<\/h2>\n
![\"$C$\"](\"2026-09-030yqdehbw50c.png\")
\u4e2a\u7c7b\u522b：<\/p>\n![\"\\\\boxed{](\"2026-09-03s0qxwlrvdm3.png\")
<\/p>\n![\"y=[0,0,1,0]\"](\"2026-09-03ln4arr2rscr.png\")
<\/p>\n![\"L=-\\\\log](\"2026-09-03mccxjmxamju.png\")
<\/p>\n
\n16. \u4e3a\u4ec0\u4e48 Cross Entropy \u4e0e Softmax \u914d\u5408？<\/h2>\n
![\"p_i=](\"2026-09-03c2vgkhpwuja.png\")
<\/p>\n![\"$z_i$\"](\"2026-09-0350o5qx54lzr.png\")
\u00a0\u662f\u6a21\u578b\u8f93\u51fa\u7684 logits\u3002<\/p>\n![\"L=-\\\\log](\"2026-09-03ubnacpr1ikt.png\")
<\/p>\n![\"L](\"2026-09-03beknau40qr4.png\")
<\/p>\n![\"\\\\boxed{](\"2026-09-033eq2t1kugkx.png\")
<\/p>\n
\n17. NLL Loss<\/h2>\n
![\"\\\\boxed{](\"2026-09-03kn2zrp2wgeg.png\")
<\/p>\n![\"\\\\boxed{](\"2026-09-03yuzwysq03lm.png\")
<\/p>\n
\n18. Cross Entropy \u4e3a\u4ec0\u4e48\u6709\u6548？<\/h2>\n
![\"y=2\"](\"2026-09-03lrzxjn2qsgh.png\")
<\/p>\n
\nclass 1: 0.1
\nclass 2: 0.8<\/p>\n![\"-\\\\log0.8](\"2026-09-035u4aydo1jxt.png\")
<\/p>\n
\nclass 1: 0.5
\nclass 2: 0.1<\/p>\n![\"-\\\\log0.1](\"2026-09-03dsdcjj1fkir.png\")
<\/p>\n
\n19. Label Smoothing<\/h2>\n
![\"y=[1,0,0]\"](\"2026-09-03oxqhxplanjt.png\")
<\/p>\n![\"y=[1-\\\\epsilon,\\\\frac{\\\\epsilon}{C-1},\\\\frac{\\\\epsilon}{C-1}]\"](\"2026-09-033mfzcqr5ls1.png\")
<\/p>\n![\"\\\\epsilon=0.1\"](\"2026-09-03yyyt0zqy2yp.png\")
<\/p>\n![\"[0.9,0.05,0.05]\"](\"2026-09-03idws5sa2oew.png\")
<\/p>\n
\n label_smoothing=0.1
\n)<\/p>\n
\n20. BCE \u4e0e Cross Entropy \u7684\u533a\u522b<\/h2>\n
\n
\u95ee\u9898BCECross Entropy<\/tr>\n \n\n \u4e8c\u5206\u7c7b<\/td>\n \u221a<\/td>\n \u221a<\/td>\n<\/tr>\n \n \u591a\u5206\u7c7b<\/td>\n \u00d7<\/td>\n \u221a<\/td>\n<\/tr>\n \n \u591a\u6807\u7b7e<\/td>\n \u221a<\/td>\n \u00d7<\/td>\n<\/tr>\n \n \u8f93\u51fa\u6fc0\u6d3b<\/td>\n Sigmoid<\/td>\n Softmax<\/td>\n<\/tr>\n \n \u7c7b\u522b\u4e4b\u95f4\u662f\u5426\u4e92\u65a5<\/td>\n \u5426<\/td>\n \u662f<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \u591a\u5206\u7c7b<\/h4>\n
\n\u72d7
\n\u9a6c<\/p>\n\u591a\u6807\u7b7e<\/h4>\n
\n\u80f8\u8154\u79ef\u6db2 = 1
\n\u6c14\u80f8 = 0<\/p>\n
\n21. KL Divergence<\/h2>\n
![\"\\\\boxed{](\"2026-09-03so4y3uhipzb.png\")
<\/p>\n
\n21.1 KL \u7684\u76f4\u89c9<\/h3>\n
![\"P=[0.7,0.2,0.1]\"](\"2026-09-03prjugg3eran.png\")
<\/p>\n![\"Q=[0.6,0.3,0.1]\"](\"2026-09-03glz245inw1b.png\")
<\/p>\n![\"Q\\\\approx](\"2026-09-03tgd3hjseeh0.png\")
<\/p>\n\n
\n22. KL \u4e0e Cross Entropy \u7684\u5173\u7cfb<\/h2>\n
![\"H(P,Q)](\"2026-09-03imyudbw4jpg.png\")
<\/p>\n![\"D_{KL}(P\\\\|Q)](\"2026-09-030tuanndmc51.png\")
<\/p>\n![\"D_{KL}(P\\\\|Q)](\"2026-09-03jmarakguinx.png\")
<\/p>\n![\"$H(P)$\"](\"2026-09-03yzjzte54zk3.png\")
\u00a0 \u4e0e\u6a21\u578b\u53c2\u6570\u65e0\u5173，\u6240\u4ee5\u5728\u4f18\u5316\u00a0![\"$Q$\"](\"2026-09-03la002k3ii3h.png\")
\u00a0\u65f6：\u6700\u5c0f\u5316 KL \u4e0e\u6700\u5c0f\u5316 Cross Entropy \u6709\u5bc6\u5207\u5173\u7cfb\u3002<\/p>\n
\n23. Hinge Loss<\/h2>\n
![\"y\\\\in\\\\{-1,+1\\\\}\"](\"2026-09-03kz4rhm4md3t.png\")
<\/p>\n![\"\\\\boxed{](\"2026-09-03frhk4t1qofi.png\")
<\/p>\n![\"yf(x)\\\\geq1\"](\"2026-09-03t3cdefigsvz.png\")
<\/p>\n![\"L=0\"](\"2026-09-03aw5e1h5bfbh.png\")
<\/p>\n![\"L>0\"](\"2026-09-03nnodtz4f4yl.png\")
<\/p>\n
\n24. Focal Loss<\/h2>\n
\n\u76ee\u6807：0.1%<\/p>\n![\"\\\\boxed{](\"2026-09-035obsprenav0.png\")
<\/p>\n![\"p_t=](\"2026-09-03biqljbhzlgg.png\")
<\/p>\n
\n25. \u4e3a\u4ec0\u4e48 Focal Loss \u6709\u6548？<\/h2>\n
![\"p_t=0.99\"](\"2026-09-03sqnsrxuzptf.png\")
<\/p>\n![\"\\\\gamma=2\"](\"2026-09-03fhmvg5qtd0l.png\")
<\/p>\n![\"(1-p_t)^2](\"2026-09-03woepnqyvkjt.png\")
<\/p>\n![\"p_t=0.2\"](\"2026-09-03atk2oehc5cb.png\")
<\/p>\n![\"(1-p_t)^2=0.64\"](\"2026-09-03lqyusb3flvx.png\")
<\/p>\n
\n26. Focal Loss PyTorch \u5b9e\u73b0<\/h2>\n
\nimport torch.nn.functional as F<\/p>\n
\n bce = F.binary_cross_entropy_with_logits(
\n logits,
\n targets.float(),
\n reduction="none"
\n )<\/p>\n
\n targets == 1,
\n probs,
\n 1 – probs
\n )<\/p>\n
\n targets == 1,
\n alpha,
\n 1 – alpha
\n )<\/p>\n
\n27. Class-Balanced Loss<\/h2>\n
![\"E_n=](\"2026-09-0304fgott5scc.png\")
<\/p>\n![\"w_n=](\"2026-09-03crxg4sgyice.png\")
<\/p>\n![\"$n$\"](\"2026-09-03xuplkyxixqf.png\")
\u662f\u8be5\u7c7b\u522b\u6837\u672c\u6570\u3002<\/p>\n![\"n\\\\downarrow](\"2026-09-03dlj22bjktnf.png\")
<\/p>\n
\n28. Dice Loss<\/h2>\n
![\"\\\\boxed{](\"2026-09-03kjinl2vkrcs.png\")
<\/p>\n\n
![\"$p_i$\"](\"2026-09-03ldl3xqvoypj.png\")
：\u9884\u6d4b\u6982\u7387<\/p>\n<\/li>\n![\"$g_i$\"](\"2026-09-03pplukia0lku.png\")
：Ground Truth<\/p>\n<\/li>\n<\/ul>\n![\"\\\\boxed{](\"2026-09-03bdswz2wzxpj.png\")
<\/p>\n![\"Dice=](\"2026-09-03stpr5gawdpm.png\")
<\/p>\n
\n29. Dice \u4e3a\u4ec0\u4e48\u7279\u522b\u9002\u5408\u5206\u5272？<\/h2>\n
\n\u75c5\u7076：10%<\/p>\n
\n\u75c5\u7076：1%<\/p>\n
\n30. Dice Loss \u7684\u4ee3\u7801<\/h2>\n
\n pred = pred.contiguous()
\n target = target.contiguous()<\/p>\n
\n 2 * intersection + smooth
\n ) \/ (
\n pred.sum() + target.sum() + smooth
\n )<\/p>\n
\n31. BCE + Dice Loss<\/h2>\n
![\"\\\\boxed{](\"2026-09-032pn141vjb2j.png\")
<\/p>\n
\n32. IoU Loss<\/h2>\n
![\"IoU=](\"2026-09-03vwynoffe10c.png\")
<\/p>\n![\"\\\\boxed{](\"2026-09-035areduzoz03.png\")
<\/p>\n![\"IoU=](\"2026-09-03vgsv5vrcihj.png\")
<\/p>\n
\n33. Dice \u4e0e IoU \u7684\u5173\u7cfb<\/h2>\n
![\"\\\\boxed{](\"2026-09-0310yp1kzfyci.png\")
<\/p>\n![\"\\\\boxed{](\"2026-09-03m5scmfrsb3p.png\")
<\/p>\n![\"IoU=0.8\"](\"2026-09-03xkutfmke0px.png\")
<\/p>\n![\"Dice=](\"2026-09-032si42sf3u3c.png\")
<\/p>\n
\n34. Tversky Loss<\/h2>\n
![\"\\\\boxed{](\"2026-09-033dqnamtlpu1.png\")
<\/p>\n![\"\\\\alpha+\\\\beta=1\"](\"2026-09-03nhbxhnchbmz.png\")
<\/p>\n![\"\\\\boxed{](\"2026-09-03dt3hiapcrde.png\")
<\/p>\n
\n34.1 \u5982\u4f55\u63a7\u5236 FP\/FN？<\/h3>\n
![\"\\\\beta>\\\\alpha\"](\"2026-09-03aqnk0d3sper.png\")
<\/p>\n![\"FN\"](\"2026-09-032l510x1rb4c.png\")
<\/p>\n![\"\\\\alpha>\\\\beta\"](\"2026-09-03izrxxhr2y2n.png\")
<\/p>\n![\"FP\"](\"2026-09-03k0atukhb3hi.png\")
<\/p>\n
\n35. Focal Tversky Loss<\/h2>\n
![\"\\\\boxed{](\"2026-09-03c0upfex2bsl.png\")
<\/p>\n
\n36. Boundary Loss<\/h2>\n
\n\u5668\u5b98\u8fb9\u754c
\n\u8840\u7ba1\u8fb9\u754c<\/p>\n![\"L_{boundary}](\"2026-09-03a20snk4ajhq.png\")
<\/p>\n![\"d_G(x)\"](\"2026-09-032pgxlrabm3c.png\")
<\/p>\n\u00a0\u5230 Ground Truth \u8fb9\u754c\u7684\u8ddd\u79bb\u3002<\/p>\n
\n37. Hausdorff Distance Loss<\/h2>\n
![\"HD(A,B)](\"2026-09-03qjh5lr4210i.png\")
<\/p>\n\n
\n38. \u76ee\u6807\u68c0\u6d4b\u4e2d\u7684 Bounding Box Loss<\/h2>\n
![\"L=](\"2026-09-03cih2sgntpkl.png\")
<\/p>\n\n
\n
\n39. IoU Bounding Box Loss<\/h2>\n
![\"A,\\\\](\"2026-09-030xrt0esorrb.png\")
<\/p>\n![\"IoU=](\"2026-09-032p1hitfw43k.png\")
<\/p>\n<\/p>\n
\nIoU = 0
\nIoU = 0<\/p>\n
\n40. GIoU Loss<\/h2>\n
![\"\\\\boxed{](\"2026-09-03ou1cqg5a2pq.png\")
<\/p>\n![\"C\"](\"2026-09-035bfcpqyp5gc.png\")
\u00a0\u662f\u540c\u65f6\u5305\u4f4f\u4e24\u4e2a Bounding Box \u7684\u6700\u5c0f\u5916\u63a5\u6846\u3002<\/p>\n![\"\\\\boxed{](\"2026-09-03owh5mz2s0o5.png\")
<\/p>\n![\"IoU=0\"](\"2026-09-03lmw3prpu3vq.png\")
<\/p>\n
\n41. DIoU Loss<\/h2>\n
![\"\\\\boxed{](\"2026-09-032dpbyzfnj25.png\")
<\/p>\n\n
![\"$\\\\rho$\"](\"2026-09-03n5xij3t5k40.png\")
：\u4e24\u4e2a\u4e2d\u5fc3\u70b9\u7684\u6b27\u6c0f\u8ddd\u79bb<\/p>\n<\/li>\n![\"$c$\"](\"2026-09-03hhi21xz1vus.png\")
：\u5916\u63a5\u6846\u5bf9\u89d2\u7ebf\u957f\u5ea6<\/p>\n<\/li>\n<\/ul>\n
\n42. CIoU Loss<\/h2>\n
\n
![\"L_{CIoU}](\"2026-09-03tj1j4ars4aq.png\")
<\/p>\n![\"v=](\"2026-09-03tvjmq0t5hqy.png\")
<\/p>\n
\n43. IoU \u7cfb\u5217\u600e\u4e48\u7406\u89e3？<\/h2>\n
\n \u2502
\n \u251c\u2500\u2500 GIoU：\u89e3\u51b3\u65e0\u91cd\u53e0\u95ee\u9898
\n \u2502
\n \u251c\u2500\u2500 DIoU：\u52a0\u5165\u4e2d\u5fc3\u8ddd\u79bb
\n \u2502
\n \u2514\u2500\u2500 CIoU：\u52a0\u5165\u5bbd\u9ad8\u6bd4<\/p>\n![\"\\\\boxed{](\"2026-09-03nonfto4mc1u.png\")
<\/p>\n
\n44. Contrastive Loss<\/h2>\n
![\"z_i,z_j\"](\"2026-09-03f152mdlsvog.png\")
<\/p>\n![\"y=](\"2026-09-03mxmw4lxetcw.png\")
<\/p>\n![\"\\\\boxed{](\"2026-09-03jkqhcwppwyt.png\")
<\/p>\n![\"D=\\\\|z_i-z_j\\\\|\"](\"2026-09-0301xpb1y0zc1.png\")
<\/p>\n
\n45. Contrastive Loss \u7684\u76f4\u89c9<\/h2>\n
<\/p>\n![\"L=D^2\"](\"2026-09-03mbxv0wax20a.png\")
<\/p>\n![\"D\\\\rightarrow0\"](\"2026-09-03flyntwpdt20.png\")
<\/p>\n<\/p>\n![\"L=\\\\max(0,m-D)^2\"](\"2026-09-033icni03wmra.png\")
<\/p>\n![\"D<m\"](\"2026-09-03ftyn0izfkfb.png\")
<\/p>\n
\n46. Triplet Loss<\/h2>\n
\nPositive
\nNegative<\/p>\n![\"a,p,n\"](\"2026-09-03dt4xvkgn1yq.png\")
<\/p>\n![\"d(a,p)<d(a,n)\"](\"2026-09-03okwra5dm2fd.png\")
<\/p>\n![\"\\\\boxed{](\"2026-09-03m44iyu4j13w.png\")
<\/p>\n![\"m\"](\"2026-09-03yoiblkslsdv.png\")
\u00a0\u662f margin\u3002<\/p>\n
\n47. Triplet Loss \u7684\u76f4\u89c9<\/h2>\n
![\"d(a,p)=0.3\"](\"2026-09-03ho2e0vil15d.png\")
，![\"d(a,n)=0.5\"](\"2026-09-03wfm3okbd0za.png\")
<\/p>\n![\"m=0.2\"](\"2026-09-03el0kqki5exr.png\")
<\/p>\n![\"L=](\"2026-09-03ndduplrgar5.png\")
<\/p>\n![\"d(a,n)=0.4\"](\"2026-09-03ed0yik3g5be.png\")
<\/p>\n![\"L=](\"2026-09-03gqsoldf1mfx.png\")
<\/p>\n
\n48. Cosine Embedding Loss<\/h2>\n
![\"x_1,x_2\"](\"2026-09-03zwry2mzkhh4.png\")
<\/p>\n![\"cos(x_1,x_2)](\"2026-09-03pusdrj2wbnr.png\")
<\/p>\n\n
\n49. Ranking Loss<\/h2>\n
![\"L=](\"2026-09-032rlru5h25ks.png\")
<\/p>\n<\/p>\n<\/p>\n![\"x_1>x_2\"](\"2026-09-03nvehq2oio4a.png\")
<\/p>\n
\n50. CTC Loss<\/h2>\n
\n
\nt1 t2 t3 t4 t5 t6<\/p>\n
\nE \u5bf9\u5e94 t3？
\nL \u5bf9\u5e94 t4\/t5？<\/p>\n
\n51. CTC \u7684\u6838\u5fc3\u601d\u60f3<\/h2>\n
![\"P(y|x)](\"2026-09-03j2x5onhpvsy.png\")
<\/p>\n\n
![\"$\\\\pi$\"](\"2026-09-03et21zlauxm3.png\")
：\u4e00\u79cd\u53ef\u80fd\u7684\u5bf9\u9f50\u8def\u5f84<\/p>\n<\/li>\n![\"$B$\"](\"2026-09-03cn2kelhdxni.png\")
：CTC collapse \u64cd\u4f5c<\/p>\n<\/li>\n：\u6700\u7ec8\u6807\u7b7e\u5e8f\u5217<\/p>\n<\/li>\n<\/ul>\n![\"\\\\boxed{](\"2026-09-03ahyp40rbsyf.png\")
<\/p>\n
\n52. VAE Loss<\/h2>\n
![\"\\\\boxed{](\"2026-09-0314dhgxlwucm.png\")
<\/p>\n
\n \u2502
\n \u251c\u2500\u2500 Reconstruction Loss
\n \u2502
\n \u2514\u2500\u2500 KL Divergence<\/p>\n
\n53. Reconstruction Loss<\/h2>\n
![\"L_{rec}](\"2026-09-03juai4ecs45e.png\")
<\/p>\n![\"\\\\hat{x}\\\\approx](\"2026-09-03ee53pwp3e3x.png\")
<\/p>\n
\n54. VAE \u4e2d\u7684 KL Loss<\/h2>\n
![\"q(z|x)\"](\"2026-09-03ug1ve5g2s2s.png\")
<\/p>\n![\"p(z)\"](\"2026-09-03tc533cm1c1d.png\")
<\/p>\n![\"p(z)=N(0,I)\"](\"2026-09-03jvyfru1aecs.png\")
<\/p>\n![\"L_{KL}](\"2026-09-03lm42rlp4gwn.png\")
<\/p>\n![\"q(z|x)](\"2026-09-03u2zsppap2oc.png\")
<\/p>\n![\"\\\\boxed{](\"2026-09-03cweppkjylbe.png\")
<\/p>\n![\"L=](\"2026-09-03th0bqni2ygh.png\")
<\/p>\n
\n55. GAN Loss<\/h2>\n
\n
![\"\\\\boxed{](\"2026-09-03lj1howiuzfa.png\")
<\/p>\n
\n56. Generator Loss<\/h2>\n
![\"L_G](\"2026-09-03fq5dejtzquy.png\")
<\/p>\n![\"\\\\boxed{](\"2026-09-03h4sdoy3sxb2.png\")
<\/p>\n
\n57. Diffusion Model Loss<\/h2>\n
![\"x_t](\"2026-09-03mybp00ie01l.png\")
<\/p>\n![\"\\\\epsilon\\\\sim](\"2026-09-03l2a5k2kmcis.png\")
<\/p>\n![\"\\\\epsilon_\\\\theta(x_t,t)\"](\"2026-09-03oeyispyemmi.png\")
<\/p>\n![\"\\\\boxed{](\"2026-09-03oyon1ifwkho.png\")
<\/p>\n
\n58. L1\/L2 Regularization<\/h2>\n
\nL1 Regularization<\/h3>\n
![\"\\\\boxed{](\"2026-09-03hmw5thn2kf4.png\")
<\/p>\n
\nL2 Regularization<\/h3>\n
![\"boxed{](\"2026-09-03um4jvhihotz.png\")
<\/p>\n![\"L=](\"2026-09-03rca5j2cbpmw.png\")
<\/p>\n
\n59. Elastic Net<\/h2>\n
![\"L_{Elastic}](\"2026-09-03wlfssl3f0ht.png\")
<\/p>\n
\nL2 \u2192 \u5e73\u6ed1
\nElastic Net \u2192 \u4e24\u8005\u7ed3\u5408<\/p>\n
\n60. \u591a\u4efb\u52a1\u5b66\u4e60 Loss<\/h2>\n
\n \u2502
\n \u251c\u2500\u2500 \u5206\u7c7b
\n \u251c\u2500\u2500 \u5206\u5272
\n \u2514\u2500\u2500 \u68c0\u6d4b<\/p>\n![\"L_{cls}\"](\"2026-09-03dknm11z02fa.png\")
，![\"L_{seg}\"](\"2026-09-03zcqrsv2tknh.png\")
\u00a0，![\"L_{det}\"](\"2026-09-03ikwhpmdbrn1.png\")
<\/p>\n![\"\\\\boxed{](\"2026-09-03bjd10kifmag.png\")
<\/p>\n
\n61. \u4e3a\u4ec0\u4e48\u9700\u8981 Loss Weight？<\/h2>\n
![\"L_{cls}=0.1\"](\"2026-09-03wav1ioose5f.png\")
<\/p>\n![\"L_{seg}=10\"](\"2026-09-03thbxu2vyfxt.png\")
<\/p>\n![\"L=10.1\"](\"2026-09-033fgxxwf4fmy.png\")
<\/p>\n![\"L=](\"2026-09-03gronsciqzyn.png\")
<\/p>\n![\"\\\\lambda_{cls}=1\"](\"2026-09-03k5iw2rhyhsn.png\")
\u00a0，![\"\\\\lambda_{seg}=0.1\"](\"2026-09-03qcrufxk1mjf.png\")
<\/p>\n
\n62. \u5e38\u89c1 Loss \u7ec4\u5408<\/h2>\n
\u5206\u7c7b<\/h4>\n
![\"L=CE\"](\"2026-09-03k1xwdwgvkaq.png\")
<\/p>\n
\n\u7c7b\u522b\u4e0d\u5e73\u8861\u5206\u7c7b<\/h4>\n
![\"L=Focal\"](\"2026-09-03fg0t1fle3zd.png\")
<\/p>\n![\"L=WeightedCE\"](\"2026-09-032jc2m2cum1z.png\")
<\/p>\n
\n\u4e8c\u5206\u7c7b<\/h4>\n
![\"L=BCE\"](\"2026-09-03i4lce1roszs.png\")
<\/p>\n
\n\u533b\u5b66\u56fe\u50cf\u5206\u5272<\/h4>\n
![\"L=Dice+BCE\"](\"2026-09-030bts21seorg.png\")
<\/p>\n
\n\u5c0f\u76ee\u6807\u5206\u5272<\/h4>\n
<\/p>\n
\n\u533b\u5b66\u8fb9\u754c\u5206\u5272<\/h4>\n
![\"L=Dice+Boundary\"](\"2026-09-03ytapbreewqp.png\")
<\/p>\n
\n\u591a\u4efb\u52a1<\/h4>\n
![\"L=](\"2026-09-03ux5youpbtsp.png\")
<\/p>\n
\n63. \u533b\u5b66\u56fe\u50cf\u5206\u5272\u4e3a\u4ec0\u4e48\u7ecf\u5e38\u4f7f\u7528 Dice + CE？<\/h2>\n
\n \u2193
\nBCE \u88ab\u80cc\u666f\u4e3b\u5bfc
\n \u2193
\n\u6a21\u578b\u503e\u5411\u4e8e\u9884\u6d4b\u80cc\u666f<\/p>\n
\n \u2193
\n\u964d\u4f4e\u7c7b\u522b\u4e0d\u5e73\u8861\u5f71\u54cd<\/p>\n![\"\\\\boxed{](\"2026-09-03nevo4ybwykl.png\")
<\/p>\n\n
\n64. Loss \u51fd\u6570\u5e94\u8be5\u5982\u4f55\u9009\u62e9？<\/h2>\n
\n
\u4efb\u52a1\u9996\u9009 Loss<\/tr>\n \n\n \u666e\u901a\u56de\u5f52<\/td>\n MSE<\/td>\n<\/tr>\n \n \u5f02\u5e38\u503c\u8f83\u591a\u7684\u56de\u5f52<\/td>\n MAE \/ Huber<\/td>\n<\/tr>\n \n \u4e8c\u5206\u7c7b<\/td>\n BCEWithLogits<\/td>\n<\/tr>\n \n \u591a\u5206\u7c7b<\/td>\n CrossEntropy<\/td>\n<\/tr>\n \n \u591a\u6807\u7b7e\u5206\u7c7b<\/td>\n BCEWithLogits<\/td>\n<\/tr>\n \n \u7c7b\u522b\u4e25\u91cd\u4e0d\u5e73\u8861<\/td>\n Focal \/ Weighted CE<\/td>\n<\/tr>\n \n \u56fe\u50cf\u5206\u5272<\/td>\n Dice + CE<\/td>\n<\/tr>\n \n \u5c0f\u76ee\u6807\u5206\u5272<\/td>\n Focal + Dice<\/td>\n<\/tr>\n \n \u533b\u5b66\u75c5\u7076\u5206\u5272<\/td>\n Dice \/ Tversky<\/td>\n<\/tr>\n \n \u8fb9\u754c\u654f\u611f\u5206\u5272<\/td>\n Dice + Boundary<\/td>\n<\/tr>\n \n \u76ee\u6807\u68c0\u6d4b\u6846\u56de\u5f52<\/td>\n IoU \/ GIoU \/ DIoU \/ CIoU<\/td>\n<\/tr>\n \n \u56fe\u50cf\u68c0\u7d22<\/td>\n Contrastive \/ Triplet<\/td>\n<\/tr>\n \n \u8868\u5f81\u5b66\u4e60<\/td>\n Contrastive<\/td>\n<\/tr>\n \n \u6392\u5e8f<\/td>\n Ranking Loss<\/td>\n<\/tr>\n \n OCR \/ ASR<\/td>\n CTC<\/td>\n<\/tr>\n \n VAE<\/td>\n Reconstruction + KL<\/td>\n<\/tr>\n \n GAN<\/td>\n Adversarial Loss<\/td>\n<\/tr>\n \n Diffusion<\/td>\n Noise Prediction MSE<\/td>\n<\/tr>\n \n \u591a\u4efb\u52a1<\/td>\n Weighted Sum<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
\n65. \u4e00\u4e2a\u975e\u5e38\u91cd\u8981\u7684\u95ee\u9898：Loss \u8d8a\u5c0f，\u6a21\u578b\u4e00\u5b9a\u8d8a\u597d\u5417？<\/h2>\n
![\"Loss\\\\downarrow\"](\"2026-09-03alvqewocojx.png\")
<\/p>\n![\"Validation\\\\](\"2026-09-033lialcbgft4.png\")
<\/p>\n![\"Train\\\\](\"2026-09-03jt4q53qd1sl.png\")
<\/p>\n![\"Validation\\\\](\"2026-09-03uty5crrpxjk.png\")
<\/p>\n
\nValidation Loss
\nEvaluation Metric<\/p>\n
\n66. \u4e3a\u4ec0\u4e48\u4e0d\u80fd\u53ea\u770b Loss？<\/h2>\n
![\"L=DiceLoss\"](\"2026-09-03f4egpe0soaa.png\")
<\/p>\n
\nIoU
\nHD95
\nPrecision
\nRecall
\nSensitivity
\nSpecificity<\/p>\n
\n67. Loss \u7684\u68af\u5ea6\u6bd4 Loss \u6570\u503c\u66f4\u52a0\u91cd\u8981<\/h2>\n
![\"\\\\nabla_\\\\theta](\"2026-09-03dqtzkiscqb0.png\")
<\/p>\n![\"\\\\theta](\"2026-09-03g00x0pgxact.png\")
<\/p>\n
\n68. \u4e3a\u4ec0\u4e48\u5f88\u591a Loss \u4f1a\u51fa\u73b0\u68af\u5ea6\u6d88\u5931？<\/h2>\n
![\"\\\\sigma(x)=\\\\frac1{1+e^{-x}}\"](\"2026-09-03tdjlvg0e4kr.png\")
<\/p>\n![\"\\\\sigma'(x)](\"2026-09-03mx2lmn1so4t.png\")
<\/p>\n![\"x\\\\rightarrow+\\\\infty\"](\"2026-09-03rjjuq5lpmuy.png\")
<\/p>\n![\"x\\\\rightarrow-\\\\infty\"](\"2026-09-03f1jpi0pls5d.png\")
<\/p>\n![\"\\\\sigma'(x)\\\\rightarrow0\"](\"2026-09-03m3g4jua5pe1.png\")
<\/p>\n
\n69. Reduction \u662f\u4ec0\u4e48\u610f\u601d？<\/h2>\n
mean<\/h4>\n
![\"L=2.5\"](\"2026-09-03hbjy5xzguzt.png\")
<\/p>\nsum<\/h4>\n
![\"L=10\"](\"2026-09-03b5s4zommnau.png\")
<\/p>\nnone<\/h4>\n
![\"L=[1,2,3,4]\"](\"2026-09-03sveh03yantm.png\")
<\/p>\n
\n70. \u4e3a\u4ec0\u4e48 reduction="none" \u5f88\u91cd\u8981？<\/h2>\n
\n logits,
\n target,
\n reduction="none"
\n)<\/p>\n
\n71. PyTorch \u5e38\u89c1 Loss API \u5bf9\u7167<\/h2>\n
\nMSELoss
\nSmoothL1Loss
\nHuberLoss<\/p>\n
\nBCEWithLogitsLoss<\/p>\n
\nNLLLoss<\/p>\n
\nHingeEmbeddingLoss
\nMultiMarginLoss
\nMultiLabelMarginLoss<\/p>\n
\nTripletMarginWithDistanceLoss<\/p>\n
\nGaussianNLLLoss<\/p>\n
\n72. \u4e00\u5f20\u56fe\u7406\u89e3\u6240\u6709 Loss<\/h2>\n
\n\u2502
\n\u251c\u2500\u2500 Regression
\n\u2502 \u251c\u2500\u2500 MSE
\n\u2502 \u251c\u2500\u2500 MAE
\n\u2502 \u251c\u2500\u2500 Smooth L1
\n\u2502 \u2514\u2500\u2500 Huber
\n\u2502
\n\u251c\u2500\u2500 Classification
\n\u2502 \u251c\u2500\u2500 BCE
\n\u2502 \u251c\u2500\u2500 Cross Entropy
\n\u2502 \u251c\u2500\u2500 NLL
\n\u2502 \u251c\u2500\u2500 Hinge
\n\u2502 \u2514\u2500\u2500 Focal
\n\u2502
\n\u251c\u2500\u2500 Distribution
\n\u2502 \u2514\u2500\u2500 KL Divergence
\n\u2502
\n\u251c\u2500\u2500 Segmentation
\n\u2502 \u251c\u2500\u2500 Dice
\n\u2502 \u251c\u2500\u2500 IoU
\n\u2502 \u251c\u2500\u2500 Tversky
\n\u2502 \u251c\u2500\u2500 Focal Tversky
\n\u2502 \u2514\u2500\u2500 Boundary
\n\u2502
\n\u251c\u2500\u2500 Object Detection
\n\u2502 \u251c\u2500\u2500 L1
\n\u2502 \u251c\u2500\u2500 Smooth L1
\n\u2502 \u251c\u2500\u2500 IoU
\n\u2502 \u251c\u2500\u2500 GIoU
\n\u2502 \u251c\u2500\u2500 DIoU
\n\u2502 \u2514\u2500\u2500 CIoU
\n\u2502
\n\u251c\u2500\u2500 Metric Learning
\n\u2502 \u251c\u2500\u2500 Contrastive
\n\u2502 \u251c\u2500\u2500 Triplet
\n\u2502 \u2514\u2500\u2500 Cosine
\n\u2502
\n\u251c\u2500\u2500 Sequence
\n\u2502 \u2514\u2500\u2500 CTC
\n\u2502
\n\u251c\u2500\u2500 Generative
\n\u2502 \u251c\u2500\u2500 VAE
\n\u2502 \u251c\u2500\u2500 GAN
\n\u2502 \u2514\u2500\u2500 Diffusion
\n\u2502
\n\u2514\u2500\u2500 Regularization
\n \u251c\u2500\u2500 L1
\n \u251c\u2500\u2500 L2
\n \u2514\u2500\u2500 Elastic Net<\/p>\n
\n73. \u6700\u91cd\u8981\u7684 Loss \u5bf9\u6bd4\u8868<\/h2>\n
\n
Loss\u6838\u5fc3\u601d\u60f3\u6700\u9002\u5408<\/tr>\n \n\n MSE<\/td>\n \u60e9\u7f5a\u5e73\u65b9\u8bef\u5dee<\/td>\n \u56de\u5f52<\/td>\n<\/tr>\n \n MAE<\/td>\n \u60e9\u7f5a\u7edd\u5bf9\u8bef\u5dee<\/td>\n \u9c81\u68d2\u56de\u5f52<\/td>\n<\/tr>\n \n Huber<\/td>\n \u5c0f\u8bef\u5dee MSE，\u5927\u8bef\u5dee MAE<\/td>\n \u9c81\u68d2\u56de\u5f52<\/td>\n<\/tr>\n \n BCE<\/td>\n \u4e8c\u5206\u7c7b\u6982\u7387\u8bef\u5dee<\/td>\n \u4e8c\u5206\u7c7b<\/td>\n<\/tr>\n \n CE<\/td>\n \u591a\u7c7b\u522b\u6982\u7387\u8bef\u5dee<\/td>\n \u591a\u5206\u7c7b<\/td>\n<\/tr>\n \n Focal<\/td>\n \u5173\u6ce8\u56f0\u96be\u6837\u672c<\/td>\n \u7c7b\u522b\u4e0d\u5e73\u8861<\/td>\n<\/tr>\n \n KL<\/td>\n \u5206\u5e03\u4e4b\u95f4\u7684\u5dee\u5f02<\/td>\n \u84b8\u998f\/VAE<\/td>\n<\/tr>\n \n Dice<\/td>\n \u533a\u57df\u91cd\u53e0<\/td>\n \u5206\u5272<\/td>\n<\/tr>\n \n IoU<\/td>\n \u533a\u57df\u4ea4\u5e76\u6bd4<\/td>\n \u5206\u5272\/\u68c0\u6d4b<\/td>\n<\/tr>\n \n Tversky<\/td>\n \u5206\u522b\u63a7\u5236 FP\/FN<\/td>\n \u533b\u5b66\u5206\u5272<\/td>\n<\/tr>\n \n GIoU<\/td>\n \u89e3\u51b3\u65e0\u91cd\u53e0\u6846<\/td>\n \u68c0\u6d4b<\/td>\n<\/tr>\n \n DIoU<\/td>\n \u52a0\u5165\u4e2d\u5fc3\u8ddd\u79bb<\/td>\n \u68c0\u6d4b<\/td>\n<\/tr>\n \n CIoU<\/td>\n \u52a0\u5165\u5bbd\u9ad8\u6bd4<\/td>\n \u68c0\u6d4b<\/td>\n<\/tr>\n \n Contrastive<\/td>\n \u6b63\u6837\u672c\u8fd1\u3001\u8d1f\u6837\u672c\u8fdc<\/td>\n \u5bf9\u6bd4\u5b66\u4e60<\/td>\n<\/tr>\n \n Triplet<\/td>\n \u76f8\u5bf9\u8ddd\u79bb\u6392\u5e8f<\/td>\n \u5ea6\u91cf\u5b66\u4e60<\/td>\n<\/tr>\n \n CTC<\/td>\n \u65e0\u9700\u663e\u5f0f\u5bf9\u9f50<\/td>\n OCR\/ASR<\/td>\n<\/tr>\n \n VAE Loss<\/td>\n \u91cd\u6784 + \u5206\u5e03\u7ea6\u675f<\/td>\n VAE<\/td>\n<\/tr>\n \n GAN Loss<\/td>\n \u5bf9\u6297\u8bad\u7ec3<\/td>\n GAN<\/td>\n<\/tr>\n \n Diffusion MSE<\/td>\n \u9884\u6d4b\u566a\u58f0<\/td>\n Diffusion<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
\n74. \u6700\u540e：\u5982\u4f55\u771f\u6b63\u7406\u89e3 Loss？<\/h2>\n
\u7b2c\u4e00\u95ee：\u5b83\u5728\u6bd4\u8f83\u4ec0\u4e48？<\/h3>\n
![\"y-\\\\hat](\"2026-09-03vl2h13xmzy3.png\")
<\/p>\n![\"Prediction](\"2026-09-030zmm2cepjqt.png\")
<\/p>\n![\"P\\\\](\"2026-09-03fupzgrmhs1y.png\")
<\/p>\n![\"d(a,p)\\\\](\"2026-09-03i30mwn3nu1v.png\")
<\/p>\n
\n\u7b2c\u4e8c\u95ee：\u5b83\u5728\u60e9\u7f5a\u4ec0\u4e48？<\/h3>\n
\n\u7b2c\u4e09\u95ee：\u5b83\u89e3\u51b3\u4e86\u4ec0\u4e48\u95ee\u9898？<\/h3>\n
\n\u2193
\n\u57fa\u7840\u56de\u5f52<\/p>\n
\n\u2193
\nMSE \u5bf9\u5f02\u5e38\u503c\u654f\u611f
\n\u2193
\n\u9c81\u68d2\u56de\u5f52<\/p>\n
\n\u2193
\nCE \u88ab\u5927\u91cf\u7b80\u5355\u6837\u672c\u652f\u914d
\n\u2193
\n\u56f0\u96be\u6837\u672c\u5b66\u4e60<\/p>\n
\n\u2193
\nCE\/BCE \u88ab\u80cc\u666f\u652f\u914d
\n\u2193
\n\u7c7b\u522b\u4e0d\u5e73\u8861\u5206\u5272<\/p>\n
\n\u2193
\nFP\/FN \u60e9\u7f5a\u4e0d\u5bf9\u79f0
\n\u2193
\n\u533b\u5b66\u5206\u5272<\/p>\n
\n\u2193
\nIoU \u65e0\u6cd5\u63cf\u8ff0\u4e2d\u5fc3\u8ddd\u79bb\u548c\u5bbd\u9ad8\u6bd4
\n\u2193
\n\u66f4\u597d\u7684 Bounding Box Regression<\/p>\n
\n75. \u4e00\u53e5\u8bdd\u8bb0\u4f4f\u6574\u4e2a Loss \u4f53\u7cfb<\/h2>\n
![\"MSE,\\\\](\"2026-09-03hbvcotzegi3.png\")
<\/p>\n![\"BCE,\\\\](\"2026-09-03f5lezplayij.png\")
<\/p>\n![\"Focal\"](\"2026-09-031jgbxq1zg30.png\")
<\/p>\n![\"Dice,\\\\](\"2026-09-03xgjqdo0sm0b.png\")
<\/p>\n![\"IoU,\\\\](\"2026-09-03ueglmvsd0os.png\")
<\/p>\n![\"Contrastive,\\\\](\"2026-09-034vhh42lm0lo.png\")
<\/p>\n![\"KL\"](\"2026-09-03rnby1xvpt2g.png\")
<\/p>\n![\"VAE,\\\\](\"2026-09-03i3ttzucrpai.png\")
<\/p>\n![\"\\\\boxed{](\"2026-09-03ny1gdhl5avz.png\")
<\/p>\n
\n\n