本文转载自：AI浩-RKD知识蒸馏实战：使用CoatNet蒸馏ResNet ，仅作留用学习！
文章目录
摘要
知识蒸馏（Knowledge Distillation），简称KD，将已经训练好的模型包含的知识(”Knowledge”)，蒸馏(“Distill”)提取到另一个模型里面去。Hinton在"Distilling the Knowledge in a Neural Network"首次提出了知识蒸馏（暗知识提取）的概念，通过引入与教师网络（Teacher network：复杂、但预测精度优越）相关的软目标（Soft-target）作为Total loss的一部分，以诱导学生网络（Student network：精简、低复杂度，更适合推理部署）的训练，实现知识迁移（Knowledge transfer）。本文使用RKD实现对模型的蒸馏。与上一篇（https://blog.csdn.net/hhhhhhhhhhwwwwwwwwww/article/details/127787791?spm=1001.2014.3001.5501）蒸馏的方法有所不同，RKD是对展平层的特征做蒸馏，蒸馏的loss分为二阶的距离损失Distance-wise Loss和三阶的角度损失Angle-wise Loss。
最终结论
先把结论说了吧！ Teacher网络使用coatnet_2，Student网络使用ResNet18。如下表
| 网络 | epochs | ACC |
|---|---|---|
| coatnet_2 | 100 | 92% |
| ResNet18 | 100 | 86% |
| ResNet18 +RKD | 100 | 90% |
在相同的条件下，加入知识蒸馏后，ResNet18的ACC上升了4个点，提升的还是很高的。如下图：
数据准备
数据使用我以前在图像分类任务中的数据集——植物幼苗数据集，先将数据集转为训练集和验证集。执行代码：
import glob
import os
import shutil
image_list=glob.glob('data1/*/*.png')
print(image_list)
file_dir='data'
if os.path.exists(file_dir):
print('true')
#os.rmdir(file_dir)
shutil.rmtree(file_dir)#删除再建立
os.makedirs(file_dir)
else:
os.makedirs(file_dir)
from sklearn.model_selection import train_test_split
trainval_files, val_files = train_test_split(image_list, test_size=0.3, random_state=42)
train_dir='train'
val_dir='val'
train_root=os.path.join(file_dir,train_dir)
val_root=os.path.join(file_dir,val_dir)
for file in trainval_files:
file_class=file.replace("\\","/").split('/')[-2]
file_name=file.replace("\\","/").split('/')[-1]
file_class=os.path.join(train_root,file_class)
if not os.path.isdir(file_class):
os.makedirs(file_class)
shutil.copy(file, file_class + '/' + file_name)
for file in val_files:
file_class=file.replace("\\","/").split('/')[-2]
file_name=file.replace("\\","/").split('/')[-1]
file_class=os.path.join(val_root,file_class)
if not os.path.isdir(file_class):
os.makedirs(file_class)
shutil.copy(file, file_class + '/' + file_name)
教师网络
教师网络选用coatnet_2，是一个比较大一点的网络了，模型的大小有200M。训练100个epoch，最好的模型在92%左右。RKD蒸馏是对展平层的蒸馏，所以我们需要对CoatNet网络做一些修改，如下：
在这里插入代码片
class CoAtNet(nn.Module):
def __init__(self, image_size, in_channels, num_blocks, channels, num_classes=1000, block_types=['C', 'C', 'T', 'T']):
super().__init__()
ih, iw = image_size
block = {
'C': MBConv, 'T': Transformer}
self.s0 = self._make_layer(
conv_3x3_bn, in_channels, channels[0], num_blocks[0], (ih // 2, iw // 2))
self.s1 = self._make_layer(
block[block_types[0]], channels[0], channels[1], num_blocks[1], (ih // 4, iw // 4))
self.s2 = self._make_layer(
block[block_types[1]], channels[1], channels[2], num_blocks[2], (ih // 8, iw // 8))
self.s3 = self._make_layer(
block[block_types[2]], channels[2], channels[3], num_blocks[3], (ih // 16, iw // 16))
self.s4 = self._make_layer(
block[block_types[3]], channels[3], channels[4], num_blocks[4], (ih // 32, iw // 32))
self.pool = nn.AvgPool2d(ih // 32, 1)
self.fc = nn.Linear(channels[-1], num_classes, bias=False)
def forward(self, x):
x = self.s0(x)
x = self.s1(x)
x = self.s2(x)
x = self.s3(x)
x = self.s4(x)
fea = self.pool(x).view(-1, x.shape[1])
x = self.fc(fea)
return fea, x
对模型的返回结果做修改，增加返回展平层的特征，这样Model的返回值有两个，一个是fea，一个是x。接下来开始编写teacher模型的train方法。
步骤
新建teacher_train.py，插入代码：
导入需要的库
import torch.optim as optim
import torch
import torch.nn as nn
import torch.nn.parallel
import torch.utils.data
import torch.utils.data.distributed
import torchvision.transforms as transforms
from torchvision import datasets
from torch.autograd import Variable
from model.coatnet import coatnet_2
import json
import os
定义训练和验证函数
def train(model, device, train_loader, optimizer, epoch):
model.train()
sum_loss = 0
total_num = len(train_loader.dataset)
print(total_num, len(train_loader))
for batch_idx, (data, target) in enumerate(train_loader):
data, target = Variable(data).to(device), Variable(target).to(device)
fea,output = model(data)
loss = criterion(output, target)
optimizer.zero_grad()
loss.backward()
optimizer.step()
print_loss = loss.data.item()
sum_loss += print_loss
if (batch_idx + 1) % 10 == 0:
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, (batch_idx + 1) * len(data), len(train_loader.dataset),
100. * (batch_idx + 1) / len(train_loader), loss.item()))
ave_loss = sum_loss / len(train_loader)
print('epoch:{},loss:{}'.format(epoch, ave_loss))
Best_ACC=0
# 验证过程
@torch.no_grad()
def val(model, device, test_loader):
global Best_ACC
model.eval()
test_loss = 0
correct = 0
total_num = len(test_loader.dataset)
print(total_num, len(test_loader))
with torch.no_grad():
for data, target in test_loader:
data, target = Variable(data).to(device), Variable(target).to(device)
fea,output = model(data)
loss = criterion(output, target)
_, pred = torch.max(output.data, 1)
correct += torch.sum(pred == target)
print_loss = loss.data.item()
test_loss += print_loss
correct = correct.data.item()
acc = correct / total_num
avgloss = test_loss / len(test_loader)
if acc > Best_ACC:
torch.save(model, file_dir + '/' + 'best.pth')
Best_ACC = acc
print('\nVal set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
avgloss, correct, len(test_loader.dataset), 100 * acc))
return acc
定义全局参数
if __name__ == '__main__':
# 创建保存模型的文件夹
file_dir = 'CoatNet'
if os.path.exists(file_dir):
print('true')
os.makedirs(file_dir, exist_ok=True)
else:
os.makedirs(file_dir)
# 设置全局参数
modellr = 1e-4
BATCH_SIZE = 16
EPOCHS = 50
DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
图像预处理与增强
# 数据预处理7
transform = transforms.Compose([
transforms.RandomRotation(10),
transforms.GaussianBlur(kernel_size=(5, 5), sigma=(0.1, 3.0)),
transforms.ColorJitter(brightness=0.5, contrast=0.5, saturation=0.5),
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.44127703, 0.4712498, 0.43714803], std=[0.18507297, 0.18050247, 0.16784933])
])
transform_test = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.44127703, 0.4712498, 0.43714803], std=[0.18507297, 0.18050247, 0.16784933])
])
读取数据
使用pytorch默认读取数据的方式。
# 读取数据
dataset_train = datasets.ImageFolder('data/train', transform=transform)
dataset_test = datasets.ImageFolder("data/val", transform=transform_test)
with open('class.txt', 'w') as file:
file.write(str(dataset_train.class_to_idx))
with open('class.json', 'w', encoding='utf-8') as file:
file.write(json.dumps(dataset_train.class_to_idx))
# 导入数据
train_loader = torch.utils.data.DataLoader(dataset_train, batch_size=BATCH_SIZE, shuffle=True)
test_loader = torch.utils.data.DataLoader(dataset_test, batch_size=BATCH_SIZE, shuffle=False)
设置模型和Loss
# 实例化模型并且移动到GPU
criterion = nn.CrossEntropyLoss()
model_ft = coatnet_2()
num_ftrs = model_ft.fc.in_features
model_ft.fc = nn.Linear(num_ftrs, 12)
model_ft.to(DEVICE)
# 选择简单暴力的Adam优化器，学习率调低
optimizer = optim.Adam(model_ft.parameters(), lr=modellr)
cosine_schedule = optim.lr_scheduler.CosineAnnealingLR(optimizer=optimizer, T_max=20, eta_min=1e-9)
# 训练
val_acc_list= {
}
for epoch in range(1, EPOCHS + 1):
train(model_ft, DEVICE, train_loader, optimizer, epoch)
cosine_schedule.step()
acc=val(model_ft, DEVICE, test_loader)
val_acc_list[epoch]=acc
with open('result.json', 'w', encoding='utf-8') as file:
file.write(json.dumps(val_acc_list))
torch.save(model_ft, 'CoatNet/model_final.pth')
完成上面的代码就可以开始训练Teacher网络了。
学生网络
学生网络选用ResNet18，是一个比较小一点的网络了，模型的大小有40M。训练50个epoch，最好的模型在86%左右。由于RKD是对展平层做蒸馏，所以还需要对ResNet做修改，使其能够返回展平层的特征。
def _forward_impl(self, x: Tensor) -> Tensor:
# See note [TorchScript super()]
x = self.conv1(x)
x = self.bn1(x)
x = self.relu(x)
x = self.maxpool(x)
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
x = self.avgpool(x)
fea = torch.flatten(x, 1)
x = self.fc(fea)
return fea,x
def forward(self, x: Tensor) -> Tensor:
return self._forward_impl(x)
返回值增加fea，然后就能获取Student模型的展平层特征了。接下来开始编写Student的训练方法。
步骤
新建student_train.py，插入代码：
导入需要的库
import torch.optim as optim
import torch
import torch.nn as nn
import torch.nn.parallel
import torch.utils.data
import torch.utils.data.distributed
import torchvision.transforms as transforms
from torchvision import datasets
from torch.autograd import Variable
from torchvision.models.resnet import resnet18
import json
import os
定义训练和验证函数
def train(model, device, train_loader, optimizer, epoch):
model.train()
sum_loss = 0
total_num = len(train_loader.dataset)
print(total_num, len(train_loader))
for batch_idx, (data, target) in enumerate(train_loader):
data, target =data.to(device), target.to(device)
fea,output = model(data)
loss = criterion(output, target)
optimizer.zero_grad()
loss.backward()
optimizer.step()
print_loss = loss.data.item()
sum_loss += print_loss
if (batch_idx + 1) % 10 == 0:
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, (batch_idx + 1) * len(data), len(train_loader.dataset),
100. * (batch_idx + 1) / len(train_loader), loss.item()))
ave_loss = sum_loss / len(train_loader)
print('epoch:{},loss:{}'.format(epoch, ave_loss))
Best_ACC=0
# 验证过程
@torch.no_grad()
def val(model, device, test_loader):
global Best_ACC
model.eval()
test_loss = 0
correct = 0
total_num = len(test_loader.dataset)
print(total_num, len(test_loader))
with torch.no_grad():
for data, target in test_loader:
data, target = data.to(device), target.to(device)
fea,output = model(data)
loss = criterion(output, target)
_, pred = torch.max(output.data, 1)
correct += torch.sum(pred == target)
print_loss = loss.data.item()
test_loss += print_loss
correct = correct.data.item()
acc = correct / total_num
avgloss = test_loss / len(test_loader)
if acc > Best_ACC:
torch.save(model, file_dir + '/' + 'best.pth')
Best_ACC = acc
print('\nVal set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
avgloss, correct, len(test_loader.dataset), 100 * acc))
return acc
定义全局参数
if __name__ == '__main__':
# 创建保存模型的文件夹
file_dir = 'resnet'
if os.path.exists(file_dir):
print('true')
os.makedirs(file_dir, exist_ok=True)
else:
os.makedirs(file_dir)
# 设置全局参数
modellr = 1e-4
BATCH_SIZE = 16
EPOCHS = 100
DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
图像预处理与增强
# 数据预处理7
transform = transforms.Compose([
transforms.RandomRotation(10),
transforms.GaussianBlur(kernel_size=(5, 5), sigma=(0.1, 3.0)),
transforms.ColorJitter(brightness=0.5, contrast=0.5, saturation=0.5),
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.44127703, 0.4712498, 0.43714803], std=[0.18507297, 0.18050247, 0.16784933])
])
transform_test = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.44127703, 0.4712498, 0.43714803], std=[0.18507297, 0.18050247, 0.16784933])
])
读取数据
使用pytorch默认读取数据的方式。
# 读取数据
dataset_train = datasets.ImageFolder('data/train', transform=transform)
dataset_test = datasets.ImageFolder("data/val", transform=transform_test)
with open('class.txt', 'w') as file:
file.write(str(dataset_train.class_to_idx))
with open('class.json', 'w', encoding='utf-8') as file:
file.write(json.dumps(dataset_train.class_to_idx))
# 导入数据
train_loader = torch.utils.data.DataLoader(dataset_train, batch_size=BATCH_SIZE, shuffle=True)
test_loader = torch.utils.data.DataLoader(dataset_test, batch_size=BATCH_SIZE, shuffle=False)
设置模型和Loss
# 实例化模型并且移动到GPU
criterion = nn.CrossEntropyLoss()
model_ft = resnet18()
print(model_ft)
num_ftrs = model_ft.fc.in_features
model_ft.fc = nn.Linear(num_ftrs, 12)
model_ft.to(DEVICE)
# 选择简单暴力的Adam优化器，学习率调低
optimizer = optim.Adam(model_ft.parameters(), lr=modellr)
cosine_schedule = optim.lr_scheduler.CosineAnnealingLR(optimizer=optimizer, T_max=20, eta_min=1e-9)
# 训练
val_acc_list= {
}
for epoch in range(1, EPOCHS + 1):
train(model_ft, DEVICE, train_loader, optimizer, epoch)
cosine_schedule.step()
acc=val(model_ft, DEVICE, test_loader)
val_acc_list[epoch]=acc
with open('result_student.json', 'w', encoding='utf-8') as file:
file.write(json.dumps(val_acc_list))
torch.save(model_ft, 'resnet/model_final.pth')
完成上面的代码就可以开始训练Student网络了。
蒸馏学生网络
学生网络继续选用ResNet18，使用Teacher网络蒸馏学生网络，训练100个epoch，最终ACC是90%。
步骤
新建student_rkd_train.py，插入代码：
导入需要的库
import torch.optim as optim
import torch
import torch.nn as nn
import torch.nn.parallel
import torch.utils.data
import torch.utils.data.distributed
import torchvision.transforms as transforms
from torchvision import datasets
from torch.autograd import Variable
from torchvision.models.resnet import resnet18
import json
import os
定义蒸馏脚本
新建rkd.py,插入代码：
from __future__ import absolute_import
from __future__ import print_function
from __future__ import division
import torch
import torch.nn as nn
import torch.nn.functional as F
'''
From https://github.com/lenscloth/RKD/blob/master/metric/loss.py
'''
class RKD(nn.Module):
'''
Relational Knowledge Distillation
https://arxiv.org/pdf/1904.05068.pdf
'''
def __init__(self, w_dist, w_angle):
super(RKD, self).__init__()
self.w_dist = w_dist
self.w_angle = w_angle
def forward(self, feat_s, feat_t):
loss = self.w_dist * self.rkd_dist(feat_s, feat_t) + \
self.w_angle * self.rkd_angle(feat_s, feat_t)
return loss
def rkd_dist(self, feat_s, feat_t):
feat_t_dist = self.pdist(feat_t, squared=False)
mean_feat_t_dist = feat_t_dist[feat_t_dist>0].mean()
feat_t_dist = feat_t_dist / mean_feat_t_dist
feat_s_dist = self.pdist(feat_s, squared=False)
mean_feat_s_dist = feat_s_dist[feat_s_dist>0].mean()
feat_s_dist = feat_s_dist / mean_feat_s_dist
loss = F.smooth_l1_loss(feat_s_dist, feat_t_dist)
return loss
def rkd_angle(self, feat_s, feat_t):
# N x C --> N x N x C
feat_t_vd = (feat_t.unsqueeze(0) - feat_t.unsqueeze(1))
norm_feat_t_vd = F.normalize(feat_t_vd, p=2, dim=2)
feat_t_angle = torch.bmm(norm_feat_t_vd, norm_feat_t_vd.transpose(1, 2)).view(-1)
feat_s_vd = (feat_s.unsqueeze(0) - feat_s.unsqueeze(1))
norm_feat_s_vd = F.normalize(feat_s_vd, p=2, dim=2)
feat_s_angle = torch.bmm(norm_feat_s_vd, norm_feat_s_vd.transpose(1, 2)).view(-1)
loss = F.smooth_l1_loss(feat_s_angle, feat_t_angle)
return loss
def pdist(self, feat, squared=False, eps=1e-12):
feat_square = feat.pow(2).sum(dim=1)
feat_prod = torch.mm(feat, feat.t())
feat_dist = (feat_square.unsqueeze(0) + feat_square.unsqueeze(1) - 2 * feat_prod).clamp(min=eps)
if not squared:
feat_dist = feat_dist.sqrt()
feat_dist = feat_dist.clone()
feat_dist[range(len(feat)), range(len(feat))] = 0
return feat_dist
定义训练和验证函数
# 定义训练过程
def train(s_net,t_net, device, criterionCls,criterionKD,train_loader, optimizer, epoch):
s_net.train()
sum_loss = 0
total_num = len(train_loader.dataset)
print(total_num, len(train_loader))
for batch_idx, (data, target) in enumerate(train_loader):
data, target = data.to(device), target.to(device)
optimizer.zero_grad()
fea_s,output = s_net(data)
cls_loss = criterionCls(output, target)
fea_t, teacher_output = t_net(data) # 训练出教师的 teacher_output
kd_loss = criterionKD(fea_s, fea_t.detach()) * lambda_kd
loss = cls_loss + kd_loss
loss.backward()
optimizer.step()
print_loss = loss.data.item()
sum_loss += print_loss
if (batch_idx + 1) % 10 == 0:
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, (batch_idx + 1) * len(data), len(train_loader.dataset),
100. * (batch_idx + 1) / len(train_loader), loss.item()))
ave_loss = sum_loss / len(train_loader)
print('epoch:{},loss:{}'.format(epoch, ave_loss))
Best_ACC=0
# 验证过程
@torch.no_grad()
def val(model, device,criterionCls, test_loader):
global Best_ACC
model.eval()
test_loss = 0
correct = 0
total_num = len(test_loader.dataset)
print(total_num, len(test_loader))
with torch.no_grad():
for data, target in test_loader:
data, target = data.to(device), target.to(device)
_,output = model(data)
loss = criterionCls(output, target)
_, pred = torch.max(output.data, 1)
correct += torch.sum(pred == target)
print_loss = loss.data.item()
test_loss += print_loss
correct = correct.data.item()
acc = correct / total_num
avgloss = test_loss / len(test_loader)
if acc > Best_ACC:
torch.save(model, file_dir + '/' + 'best.pth')
Best_ACC = acc
print('\nVal set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
avgloss, correct, len(test_loader.dataset), 100 * acc))
return acc
定义全局参数
if __name__ == '__main__':
# 创建保存模型的文件夹
file_dir = 'resnet_rkd'
if os.path.exists(file_dir):
print('true')
os.makedirs(file_dir, exist_ok=True)
else:
os.makedirs(file_dir)
# 设置全局参数
modellr = 1e-4
BATCH_SIZE = 16
EPOCHS = 100
DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
w_dist=25
w_angle=50
lambda_kd=1.0
图像预处理与增强
# 数据预处理7
transform = transforms.Compose([
transforms.RandomRotation(10),
transforms.GaussianBlur(kernel_size=(5, 5), sigma=(0.1, 3.0)),
transforms.ColorJitter(brightness=0.5, contrast=0.5, saturation=0.5),
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.44127703, 0.4712498, 0.43714803], std=[0.18507297, 0.18050247, 0.16784933])
])
transform_test = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.44127703, 0.4712498, 0.43714803], std=[0.18507297, 0.18050247, 0.16784933])
])
读取数据
使用pytorch默认读取数据的方式。
# 读取数据
dataset_train = datasets.ImageFolder('data/train', transform=transform)
dataset_test = datasets.ImageFolder("data/val", transform=transform_test)
with open('class.txt', 'w') as file:
file.write(str(dataset_train.class_to_idx))
with open('class.json', 'w', encoding='utf-8') as file:
file.write(json.dumps(dataset_train.class_to_idx))
# 导入数据
train_loader = torch.utils.data.DataLoader(dataset_train, batch_size=BATCH_SIZE, shuffle=True)
test_loader = torch.utils.data.DataLoader(dataset_test, batch_size=BATCH_SIZE, shuffle=False)
设置模型和Loss
# 实例化模型并且移动到GPU
model_ft = resnet18()
print(model_ft)
num_ftrs = model_ft.fc.in_features
model_ft.fc = nn.Linear(num_ftrs, 12)
model_ft.to(DEVICE)
# 选择简单暴力的Adam优化器，学习率调低
optimizer = optim.Adam(model_ft.parameters(), lr=modellr)
cosine_schedule = optim.lr_scheduler.CosineAnnealingLR(optimizer=optimizer, T_max=20, eta_min=1e-9)
teacher_model=torch.load('./CoatNet/best.pth')
teacher_model.eval()
# 定义rkd loss和nn.CrossEntropyLoss()
criterionKD = RKD(w_dist, w_angle)
criterionCls = nn.CrossEntropyLoss()
# 训练
val_acc_list= {
}
for epoch in range(1, EPOCHS + 1):
train(model_ft,teacher_model, DEVICE,criterionCls,criterionKD, train_loader, optimizer, epoch)
cosine_schedule.step()
acc=val(model_ft,DEVICE,criterionCls , test_loader)
val_acc_list[epoch]=acc
with open('result_rkd.json', 'w', encoding='utf-8') as file:
file.write(json.dumps(val_acc_list))
torch.save(model_ft, 'resnet_rkd/model_final.pth')
完成上面的代码就可以开始蒸馏模式！！！
结果比对
加载保存的结果，然后绘制acc曲线。
import numpy as np
from matplotlib import pyplot as plt
import json
teacher_file='result.json'
student_file='result_student.json'
student_kd_file='result_rkd.json'
def read_json(file):
with open(file, 'r', encoding='utf8') as fp:
json_data = json.load(fp)
print(json_data)
return json_data
teacher_data=read_json(teacher_file)
student_data=read_json(student_file)
student_kd_data=read_json(student_kd_file)
x =[int(x) for x in list(dict(teacher_data).keys())]
print(x)
plt.plot(x, list(teacher_data.values()), label='teacher')
plt.plot(x,list(student_data.values()), label='student without RKD')
plt.plot(x, list(student_kd_data.values()), label='student with RKD')
plt.title('Test accuracy')
plt.legend()
plt.show()
总结
知识蒸馏是常用的一种对轻量化模型压缩和提升的方法。今天通过一个简单的例子讲解了如何使用Teacher网络对Student网络进行RKD蒸馏。
扫一扫
1072
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
