手动实现ArcFaceLoss和CenterLoss，并用来训练MNIST数据。
导入相关库
import torch
import torch.nn as nn
import torch.optim as optim
from torchvision import datasets, transforms as T
from torch.utils.data import DataLoader
import itertools
import matplotlib.pyplot as plt
# 查看时间和进度
from tqdm import tqdm
import time
实现ArcFaceNet和CenterLossNet
- ArcFaceLoss参考：
- 论文地址：https://arxiv.org/abs/1801.07698
- 代码地址：https://github.com/deepinsight/insightface
- CenterLoss参考：
- 论文地址：https://ydwen.github.io/papers/WenECCV16.pdf
- 代码地址：https://github.com/ydwen/caffe-face
class ArcFaceNet(nn.Module):
def __init__(self, cls_num=10, feature_dim=2):
super(ArcFaceNet, self).__init__()
self.w = nn.Parameter(torch.randn(feature_dim, cls_num))
def forward(self, features, m=1, s=10):
# 特征与权重 归一化
_features = nn.functional.normalize(features, dim=1)
_w = nn.functional.normalize(self.w, dim=0)
# 特征向量与参数向量的夹角theta，分子numerator，分母denominator
theta = torch.acos(torch.matmul(_features, _w) / 10) # /10防止下溢
numerator = torch.exp(s * torch.cos(theta + m))
denominator = torch.sum(torch.exp(s * torch.cos(theta)), dim=1, keepdim=True) - torch.exp(
s * torch.cos(theta)) + numerator
return torch.log(torch.div(numerator, denominator))
class CenterLossNet(nn.Module):
def __init__(self, cls_num=10, feature_dim=2):
super(CenterLossNet, self).__init__()
self.centers = nn.Parameter(torch.randn(cls_num, feature_dim))
def forward(self, features, labels, reduction='mean'):
# 特征向量归一化
_features = nn.functional.normalize(features)
centers_batch = self.centers.index_select(dim=0, index=labels.long())
# 根据论文《A Discriminative Feature Learning Approach for Deep Face Recognition》修改如下
if reduction == 'sum': # 返回loss的和
return torch.sum(torch.pow(_features - centers_batch, 2)) / 2
elif reduction == 'mean': # 返回loss和的平均值，默认为mean方式
return torch.sum(torch.pow(_features - centers_batch, 2)) / 2 / len(features)
else:
raise ValueError("ValueError: {0} is not a valid value for reduction".format(reduction))
定义LeNet模型
class LeNet(nn.Module):
def __init__(self):
super().__init__()
self.conv = nn.Sequential(
nn.Conv2d(1, 64, 3, padding=1),
nn.PReLU(),
nn.BatchNorm2d(64),
nn.Conv2d(64, 32, 3, stride=2, padding=1),
nn.PReLU(),
nn.BatchNorm2d(32),
nn.modules.Flatten()
)
self.linear = nn.Sequential(
nn.Linear(32 * 14 * 14, 512),
nn.PReLU(),
nn.BatchNorm1d(512),
nn.Linear(512, 256),
nn.PReLU(),
nn.BatchNorm1d(256),
nn.Linear(256, 64),
nn.PReLU(),
nn.BatchNorm1d(64),
nn.Linear(64, 32)
# nn.Linear(64, 2, bias=False) # features设置为二维，可以进行可视化
)
self.out_layer = nn.Sequential(
nn.Linear(32, 10),
# nn.Linear(2, 10), # features设置为二维，可以进行可视化
nn.LogSoftmax(dim=1) # LogSoftmax与net=nn.NLLLoss()结合使用,求交叉熵损失
)
def forward(self, x):
features = self.linear(self.conv(x))
out = self.out_layer(features) # 用于计算CrossEntropyLoss
return features, out
模型训练
两种损失计算方式：
- CrossEntropyLoss+CenteLoss
- ArcFaceLoss+CenterLoss
超参数都是初始随便设定的，跑了一遍，精度可达到99.29。你可以调调超参数，精度可以更高。训练代码如下：
# 特征向量可视化
def visualize(features, labels, loss, epoch):
# 定义10种颜色
colors = ['#ff0000', '#ffff00', '#00ff00', '#00ffff', '#0000ff', '#ff00ff', '#990000', '#999900', '#009900',
'#009999']
plt.clf() # 清空画板
# 画出所有的点，不同的label对应不同的颜色
for i in range(10):
plt.plot(features[labels == i, 0], features[labels == i, 1], ".", c=colors[i], label=i)
plt.legend(loc="upper right") # 图例
plt.title(f"ce+cl: epoch={epoch}, loss={loss}") # 标题
plt.savefig("ce+cl/image/epoch%d.jpg" % epoch) # 保存图片
plt.draw() # 展示图片
plt.pause(0.001)
# 1.加载数据集
transform_op = T.Compose([ # 数据预处理
T.ToTensor(),
T.Normalize([0.4914], [0.2023])
])
train_dataset = datasets.MNIST("../code/data", train=True, transform=transform_op, download=False)
val_dataset = datasets.MNIST("../code/data", train=False, transform=transform_op, download=False)
train_dataloader = DataLoader(train_dataset, batch_size=128, shuffle=True)
val_dataloader = DataLoader(val_dataset, batch_size=128, shuffle=False)
# 训练设备: GPU or CPU
device= torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# 超参
lr = 1e-3
epochs = 20
lr_schedule = {
5: 1e-3,
10: 1e-4,
15: 1e-5
}
alpha = 0.95 # centerloss与arcfaceloss的权重比例
1.CrossEntropyLoss+CenteLoss
# 2.创建模型
cls_num, feature_dim = 10, 32 # 10分类
# cls_num, feature_dim = 10, 2 # features设置为二维，可以进行可视化
net = LeNet().to(device)
centerloss_net = CenterLossNet(cls_num, feature_dim).to(device)
# 3.定义损失
loss_func = nn.NLLLoss()
# 4.定义优化器
optimizer = optim.Adam(itertools.chain(net.parameters(), centerloss_net.parameters()), lr)
# 5.模型训练
plt.ion()
for epoch in range(epochs):
start = time.time()
# 学习率策略
if epoch in lr_schedule:
lr = lr_schedule[epoch]
for group in optimizer.param_groups:
group["lr"] = lr
# 1）训练集
net.train() # train mode
features_loader, labels_loader = [], [] # 保存特征向量和标签的列表，用于可视化操作
train_loss = 0.
for images, targets in tqdm(train_dataloader):
images, targets = images.to(device), targets.to(device)
# 方式1: CrossEntropyLoss+CenterLoss
features, out = net(images)
# 计算损失
ce_loss = loss_func(out, targets)
center_loss = centerloss_net(features, targets)
loss = alpha * ce_loss + (1 - alpha) * center_loss
optimizer.zero_grad() # 清空梯度
loss.backward() # 反向传播
optimizer.step() # 梯度更新
# 统计训练损失
train_loss += loss.cpu().detach().item()
# 将特征和标签加入到列表中
features_loader.append(features)
labels_loader.append(targets)
# 计算平均损失
train_loss /= len(train_dataloader)
# 2.测试集
net.eval() # evaluation mode
val_loss, correct = 0., 0.
with torch.no_grad(): # 作用域范围内不计算梯度，节省内存
for images, targets in tqdm(val_dataloader):
images, targets = images.to(device), targets.to(device)
# 方式1: CrossEntropyLoss+CenterLoss
features, out = net(images)
# 计算损失
ce_loss = loss_func(out, targets)
center_loss = centerloss_net(features, targets)
loss = alpha * ce_loss + (1 - alpha) * center_loss
# 统计验证损失
val_loss += loss.cpu().detach().item()
# 统计正确的个数
correct += sum(out.argmax(1) == targets)
# 计算平均损失
val_loss /= len(val_dataloader)
# 计算准确率
accuracy = correct.item() / len(val_dataset)
# 打印损失和精度信息
print(f"Epoch: {epoch}/{epochs}, Train_loss: {train_loss:.5f}, Val_loss: {val_loss:.5f}, Accuracy: {accuracy}")
# 保存模型参数
torch.save(net.state_dict(), f"ce+cl/checkpoint/net.pt")
torch.save(centerloss_net.state_dict(), f"ce+cl/checkpoint/centerloss_net.pt")
# 特征向量可视化
features = torch.cat(features_loader, dim=0)
labels = torch.cat(labels_loader, dim=0)
visualize(features.cpu().detach().numpy(), labels.cpu().detach().numpy(), train_loss, epoch)
# 查看时间和进度
end = time.time() # 本次轮询结束时间
print(f"第{epoch}次轮询，共耗时{int(end - start)}秒")
time.sleep(0.01)
plt.ioff()
2.ArcFaceLoss+CenterLoss
# 2.创建模型
cls_num, feature_dim = 10, 32 # 10分类
# cls_num, feature_dim = 10, 2 # features设置为二维，可以进行可视化
net = LeNet().to(device)
arcface_net = ArcFaceNet(cls_num, feature_dim).to(device)
centerloss_net = CenterLossNet(cls_num, feature_dim).to(device)
# 3.定义损失
loss_func = nn.NLLLoss()
# 4.定义优化器
optimizer = optim.Adam(itertools.chain(net.parameters(), arcface_net.parameters(), centerloss_net.parameters()), lr)
# 5.模型训练
plt.ion()
for epoch in range(epochs):
start = time.time()
# 学习率策略
if epoch in lr_schedule:
lr = lr_schedule[epoch]
for group in optimizer.param_groups:
group["lr"] = lr
# 1）训练集
net.train() # train mode
features_loader, labels_loader = [], [] # 保存特征向量和标签的列表，用于可视化操作
train_loss = 0.
for images, targets in tqdm(train_dataloader):
images, targets = images.to(device), targets.to(device)
# 方式2: ArcFaceLoss+CenterLoss
features, _ = net(images)
out = arcface_net(features)
# 计算损失
arcface_loss = loss_func(out, targets) # arcfaceloss
center_loss = centerloss_net(features, targets) # centerloss
loss = alpha * arcface_loss + (1 - alpha) * center_loss
optimizer.zero_grad() # 清空梯度
loss.backward() # 反向传播
optimizer.step() # 梯度更新
# 统计训练损失
train_loss += loss.cpu().detach().item()
# 将特征和标签加入到列表中
features_loader.append(features)
labels_loader.append(targets)
# 计算平均损失
train_loss /= len(train_dataloader)
# 2.测试集
net.eval() # evaluation mode
val_loss, correct = 0., 0.
with torch.no_grad(): # 作用域范围内不计算梯度，节省内存
for images, targets in tqdm(val_dataloader):
images, targets = images.to(device), targets.to(device)
# 方式2: ArcFaceLoss+CenterLoss
features, _ = net(images)
out = arcface_net(features)
# 计算损失
arcface_loss = loss_func(out, targets) # arcfaceloss
center_loss = centerloss_net(features, targets) # centerloss
loss = alpha * arcface_loss + (1 - alpha) * center_loss
# 统计验证损失
val_loss += loss.cpu().detach().item()
# 统计正确的个数
correct += sum(out.argmax(1) == targets)
# 计算平均损失
val_loss /= len(val_dataloader)
# 计算准确率
accuracy = correct.item() / len(val_dataset)
# 打印损失和精度信息
print(alpha * arcface_loss, (1 - alpha) * center_loss, arcface_loss, center_loss)
print(f"Epoch: {epoch}/{epochs}, Train_loss: {train_loss:.5f}, Val_loss: {val_loss:.5f}, Accuracy: {accuracy}")
# 保存模型参数
torch.save(net.state_dict(), f"arcface+cl/checkpoint/net.pt")
torch.save(centerloss_net.state_dict(), f"arcface+cl/checkpoint/centerloss_net.pt")
torch.save(arcface_net.state_dict(), f"arcface+cl/checkpoint/arcface_net.pt")
# 特征向量可视化
features = torch.cat(features_loader, dim=0)
labels = torch.cat(labels_loader, dim=0)
visualize(features.cpu().detach().numpy(), labels.cpu().detach().numpy(), epoch, train_loss, val_loss, accuracy)
# 查看时间和进度
end = time.time() # 本次轮询结束时间
print(f"第{epoch}次轮询，共耗时{int(end - start)}秒")
time.sleep(0.01)
plt.ioff()