迁移学习（五）：知识蒸馏

  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import DataLoader
import torchvision
import torchvision.transforms as transforms
from typing import List
import copy

# ===== 损失函数 =====

class KDLoss(nn.Module):
    """响应蒸馏：带温度的 KL 散度，混合硬标签交叉熵。"""
    def __init__(self, temperature: float = 4.0, alpha: float = 0.9):
        super().__init__()
        self.T = temperature
        self.alpha = alpha

    def forward(self, student_logits, teacher_logits, labels):
        kd = F.kl_div(
            F.log_softmax(student_logits / self.T, dim=1),
            F.softmax(teacher_logits / self.T, dim=1),
            reduction='batchmean',
        ) * (self.T ** 2)
        ce = F.cross_entropy(student_logits, labels)
        return self.alpha * kd + (1 - self.alpha) * ce

class FeatureDistillLoss(nn.Module):
    """FitNets：1x1 卷积投影后计算特征图均方误差。"""
    def __init__(self, student_ch: int, teacher_ch: int):
        super().__init__()
        self.proj = nn.Conv2d(student_ch, teacher_ch, 1, bias=False)

    def forward(self, student_feat, teacher_feat):
        return F.mse_loss(self.proj(student_feat), teacher_feat)

class AttentionTransferLoss(nn.Module):
    """匹配通道聚合后的归一化空间注意力图。"""
    def __init__(self, p: float = 2.0):
        super().__init__()
        self.p = p

    def _attn(self, feat):
        a = torch.sum(torch.abs(feat) ** self.p, dim=1, keepdim=True)
        return a / (a.sum(dim=[2, 3], keepdim=True) + 1e-8)

    def forward(self, s_feat, t_feat):
        return F.mse_loss(self._attn(s_feat), self._attn(t_feat))

class RelationalLoss(nn.Module):
    """RKD：样本间的距离和角度关系。"""
    def __init__(self, w_dist: float = 1.0, w_angle: float = 2.0):
        super().__init__()
        self.w_dist, self.w_angle = w_dist, w_angle

    def forward(self, s_feat, t_feat):
        s = F.normalize(s_feat, p=2, dim=1)
        t = F.normalize(t_feat, p=2, dim=1)
        d_loss = F.mse_loss(torch.cdist(s, s), torch.cdist(t, t))
        a_loss = F.mse_loss(s @ s.t(), t @ t.t())
        return self.w_dist * d_loss + self.w_angle * a_loss

# ===== 模型（返回 logits 和中间特征）=====

class _ResNetBackbone(nn.Module):
    def __init__(self, ctor, num_classes: int):
        super().__init__()
        self.net = ctor(weights=None)
        self.net.fc = nn.Linear(self.net.fc.in_features, num_classes)

    def forward(self, x):
        n = self.net
        x = n.maxpool(n.relu(n.bn1(n.conv1(x))))
        feats: List[torch.Tensor] = []
        for layer in (n.layer1, n.layer2, n.layer3, n.layer4):
            x = layer(x)
            feats.append(x)
        logits = n.fc(torch.flatten(n.avgpool(x), 1))
        return logits, feats

class TeacherNet(_ResNetBackbone):
    def __init__(self, num_classes: int = 10):
        super().__init__(torchvision.models.resnet34, num_classes)

class StudentNet(_ResNetBackbone):
    def __init__(self, num_classes: int = 10):
        super().__init__(torchvision.models.resnet18, num_classes)

# ===== 训练器（支持多种蒸馏模式）=====

class DistillationTrainer:
    def __init__(self, teacher, student, device='cpu',
                 mode: str = 'response',
                 temperature: float = 4.0, alpha: float = 0.9):
        self.teacher = teacher.to(device).eval()
        self.student = student.to(device)
        self.device = device
        self.mode = mode
        for p in self.teacher.parameters():
            p.requires_grad = False

        self.kd = KDLoss(temperature, alpha)
        if mode in ('feature', 'combined'):
            ch = [64, 128, 256, 512]
            self.feat = nn.ModuleList(
                [FeatureDistillLoss(s, t).to(device)
                 for s, t in zip(ch, ch)])
        if mode in ('attention', 'combined'):
            self.attn = AttentionTransferLoss()
        if mode in ('relation', 'combined'):
            self.rel = RelationalLoss()

    def _loss(self, s_logits, s_feats, t_logits, t_feats, y):
        loss = self.kd(s_logits, t_logits, y)
        if self.mode in ('feature', 'combined'):
            fl = sum(fn(sf, tf)
                     for fn, sf, tf in zip(self.feat, s_feats, t_feats))
            loss = loss + 0.5 * fl / len(s_feats)
        if self.mode in ('attention', 'combined'):
            al = sum(self.attn(sf, tf)
                     for sf, tf in zip(s_feats, t_feats))
            loss = loss + 0.3 * al / len(s_feats)
        if self.mode in ('relation', 'combined'):
            loss = loss + 0.2 * self.rel(
                s_feats[-1].flatten(1), t_feats[-1].flatten(1))
        return loss

    def train_epoch(self, loader, optimizer):
        self.student.train()
        total, correct, n = 0.0, 0, 0
        for x, y in loader:
            x, y = x.to(self.device), y.to(self.device)
            with torch.no_grad():
                t_logits, t_feats = self.teacher(x)
            s_logits, s_feats = self.student(x)
            loss = self._loss(s_logits, s_feats, t_logits, t_feats, y)
            optimizer.zero_grad(); loss.backward(); optimizer.step()
            total += loss.item() * y.size(0)
            correct += (s_logits.argmax(1) == y).sum().item()
            n += y.size(0)
        return total / n, 100.0 * correct / n

    @torch.no_grad()
    def evaluate(self, loader):
        self.student.eval()
        correct, n = 0, 0
        for x, y in loader:
            x, y = x.to(self.device), y.to(self.device)
            logits, _ = self.student(x)
            correct += (logits.argmax(1) == y).sum().item()
            n += y.size(0)
        return 100.0 * correct / n

# ===== 自蒸馏：Born-Again Networks =====

def self_distill(model_class, train_loader, test_loader,
                 num_generations: int = 3, epochs_per_gen: int = 10,
                 device: str = 'cpu', temperature: float = 4.0):
    teacher = None
    for gen in range(num_generations):
        student = model_class().to(device)
        opt = optim.SGD(student.parameters(), lr=0.1,
                        momentum=0.9, weight_decay=5e-4)
        sched = optim.lr_scheduler.CosineAnnealingLR(opt, epochs_per_gen)

        for _ in range(epochs_per_gen):
            student.train()
            for x, y in train_loader:
                x, y = x.to(device), y.to(device)
                logits, _ = student(x)
                loss = F.cross_entropy(logits, y)
                if teacher is not None:
                    with torch.no_grad():
                        t_logits, _ = teacher(x)
                    kd = F.kl_div(
                        F.log_softmax(logits / temperature, dim=1),
                        F.softmax(t_logits / temperature, dim=1),
                        reduction='batchmean') * temperature ** 2
                    loss = 0.1 * loss + 0.9 * kd
                opt.zero_grad(); loss.backward(); opt.step()
            sched.step()

        student.eval()
        correct, n = 0, 0
        with torch.no_grad():
            for x, y in test_loader:
                x, y = x.to(device), y.to(device)
                correct += (student(x)[0].argmax(1) == y).sum().item()
                n += y.size(0)
        print(f"第 {gen + 1} 代：准确率 {100.0 * correct / n:.2f}%")

        teacher = copy.deepcopy(student).eval()
        for p in teacher.parameters():
            p.requires_grad = False

# ===== 主流程 =====

def main():
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
    norm = transforms.Normalize((0.4914, 0.4822, 0.4465),
                                (0.2023, 0.1994, 0.2010))
    train_tf = transforms.Compose([
        transforms.RandomCrop(32, padding=4),
        transforms.RandomHorizontalFlip(),
        transforms.ToTensor(), norm])
    test_tf = transforms.Compose([transforms.ToTensor(), norm])

    train_set = torchvision.datasets.CIFAR10(
        './data', train=True, download=True, transform=train_tf)
    test_set = torchvision.datasets.CIFAR10(
        './data', train=False, download=True, transform=test_tf)
    train_loader = DataLoader(train_set, 128, shuffle=True, num_workers=2)
    test_loader = DataLoader(test_set, 128, num_workers=2)

    teacher = TeacherNet(10)
    student = StudentNet(10)
    trainer = DistillationTrainer(
        teacher, student, device, mode='combined',
        temperature=4.0, alpha=0.7)
    opt = optim.SGD(student.parameters(), lr=0.1,
                    momentum=0.9, weight_decay=5e-4)
    for epoch in range(20):
        loss, acc = trainer.train_epoch(train_loader, opt)
        test_acc = trainer.evaluate(test_loader)
        print(f"第 {epoch + 1} 轮：损失 {loss:.4f} "
              f"训练准确率 {acc:.1f}% 测试准确率 {test_acc:.1f}%")

    print("\n自蒸馏（Born-Again Networks）：")
    self_distill(StudentNet, train_loader, test_loader,
                 num_generations=3, epochs_per_gen=10, device=device)

if __name__ == '__main__':
    main()

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32

import torch.nn.functional as F

class TemperatureSchedule:
    """在 T 个 epoch 内从 tau_start 线性下降到 tau_end"""
    def __init__(self, tau_start: float, tau_end: float, T: int):
        self.t0, self.t1, self.T = tau_start, tau_end, T

    def __call__(self, epoch: int) -> float:
        frac = min(epoch / max(self.T - 1, 1), 1.0)
        return self.t0 - (self.t0 - self.t1) * frac

def kd_loss_scheduled(s_logits, t_logits, y, tau: float, alpha: float = 0.9):
    kd = F.kl_div(
        F.log_softmax(s_logits / tau, dim=1),
        F.softmax(t_logits / tau, dim=1),
        reduction='batchmean',
    ) * (tau ** 2)
    ce = F.cross_entropy(s_logits, y)
    return alpha * kd + (1 - alpha) * ce

# 基准实验：CIFAR-10 上 ResNet-18 教师 → ResNet-8 学生
sched = TemperatureSchedule(tau_start=20.0, tau_end=2.0, T=epochs)
for epoch in range(epochs):
    tau = sched(epoch)
    student.train()
    for x, y in train_loader:
        x, y = x.to(device), y.to(device)
        with torch.no_grad():
            t_logits, _ = teacher(x)
        s_logits, _ = student(x)
        loss = kd_loss_scheduled(s_logits, t_logits, y, tau=tau, alpha=0.9)
        opt.zero_grad(); loss.backward(); opt.step()

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34

import torch
import torch.nn as nn
import torch.nn.functional as F

class ProjectionHead(nn.Module):
    def __init__(self, in_dim: int, out_dim: int = 128):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(in_dim, in_dim), nn.ReLU(inplace=True),
            nn.Linear(in_dim, out_dim))

    def forward(self, x):
        return F.normalize(self.net(x), p=2, dim=1)

class CRDLoss(nn.Module):
    """基于 batch 内负样本的 (student, teacher) 对 InfoNCE 损失"""
    def __init__(self, s_dim: int, t_dim: int,
                 proj_dim: int = 128, tau: float = 0.07):
        super().__init__()
        self.g_s = ProjectionHead(s_dim, proj_dim)
        self.g_t = ProjectionHead(t_dim, proj_dim)
        self.tau = tau

    def forward(self, s_feat, t_feat):
        # 如有必要，对空间维度做池化
        if s_feat.dim() == 4:
            s_feat = F.adaptive_avg_pool2d(s_feat, 1).flatten(1)
        if t_feat.dim() == 4:
            t_feat = F.adaptive_avg_pool2d(t_feat, 1).flatten(1)
        z_s = self.g_s(s_feat)             # (B, D)
        z_t = self.g_t(t_feat).detach()    # 固定教师
        logits = z_s @ z_t.t() / self.tau  # (B, B)：第 i 行第 i 列为正样本
        labels = torch.arange(z_s.size(0), device=z_s.device)
        return F.cross_entropy(logits, labels)

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40

@torch.no_grad()
def kd_diagnostic(student, teacher, loader, n_epochs: int,
                  optimizer, device='cpu', tau: float = 4.0):
    """返回每 epoch 的平均 KL(p_t || p_s)，并判断主要瓶颈"""
    teacher.eval()
    kl_history = []
    for epoch in range(n_epochs):
        student.train()
        for x, y in loader:
            x, y = x.to(device), y.to(device)
            t_logits, _ = teacher(x)
            s_logits, _ = student(x)
            loss = F.kl_div(
                F.log_softmax(s_logits / tau, dim=1),
                F.softmax(t_logits / tau, dim=1),
                reduction='batchmean') * (tau ** 2)
            optimizer.zero_grad(); loss.backward(); optimizer.step()
        # 在干净 pass 上测量 KL
        student.eval()
        kl_sum, n = 0.0, 0
        for x, _ in loader:
            x = x.to(device)
            t_logits, _ = teacher(x)
            s_logits, _ = student(x)
            p_t = F.softmax(t_logits / tau, dim=1)
            log_p_s = F.log_softmax(s_logits / tau, dim=1)
            kl_sum += F.kl_div(log_p_s, p_t, reduction='batchmean').item() * x.size(0)
            n += x.size(0)
        kl_history.append(kl_sum / n)

    init, mid, last = kl_history[0], kl_history[len(kl_history) // 2], kl_history[-1]
    if last > 0.5 * init:
        verdict = 'capacity'   # KL 在整个训练中未减半
    elif max(kl_history[-5:]) - min(kl_history[-5:]) > 0.2 * last:
        verdict = 'lr / instability'
    elif init > 5.0:
        verdict = 'data / modality mismatch'
    else:
        verdict = 'healthy'
    return kl_history, verdict