迁移学习（三）：域适应

  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

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import DataLoader, TensorDataset
from torch.autograd import Function
import numpy as np
from sklearn.metrics import accuracy_score


class GradientReversalFunction(Function):
    """前向恒等，反向把梯度取反。"""

    @staticmethod
    def forward(ctx, x, lambda_):
        ctx.lambda_ = lambda_
        return x.clone()

    @staticmethod
    def backward(ctx, grad_output):
        return grad_output.neg() * ctx.lambda_, None


class GradientReversalLayer(nn.Module):
    def __init__(self):
        super().__init__()
        self.lambda_ = 1.0

    def set_lambda(self, val):
        self.lambda_ = val

    def forward(self, x):
        return GradientReversalFunction.apply(x, self.lambda_)


class FeatureExtractor(nn.Module):
    def __init__(self, input_dim=28 * 28, hidden_dim=256):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(input_dim, hidden_dim), nn.ReLU(), nn.Dropout(0.5),
            nn.Linear(hidden_dim, hidden_dim), nn.ReLU(), nn.Dropout(0.5),
        )

    def forward(self, x):
        return self.net(x.view(x.size(0), -1))


class LabelPredictor(nn.Module):
    def __init__(self, feature_dim=256, num_classes=10):
        super().__init__()
        self.fc = nn.Linear(feature_dim, num_classes)

    def forward(self, x):
        return self.fc(x)


class DomainDiscriminator(nn.Module):
    def __init__(self, feature_dim=256):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(feature_dim, 256), nn.ReLU(), nn.Dropout(0.5),
            nn.Linear(256, 256), nn.ReLU(), nn.Dropout(0.5),
            nn.Linear(256, 1), nn.Sigmoid(),
        )

    def forward(self, x):
        return self.net(x)


class DANN(nn.Module):
    """域对抗神经网络。"""

    def __init__(self, input_dim=28 * 28, hidden_dim=256, num_classes=10):
        super().__init__()
        self.feature_extractor = FeatureExtractor(input_dim, hidden_dim)
        self.label_predictor = LabelPredictor(hidden_dim, num_classes)
        self.domain_discriminator = DomainDiscriminator(hidden_dim)
        self.grl = GradientReversalLayer()

    def forward(self, x, alpha=1.0):
        features = self.feature_extractor(x)
        class_logits = self.label_predictor(features)
        self.grl.set_lambda(alpha)
        domain_logits = self.domain_discriminator(self.grl(features))
        return class_logits, domain_logits


class DANNTrainer:
    def __init__(self, model, source_loader, target_loader, test_loader,
                 num_epochs=100, lr=1e-3, device="cpu", gamma=10.0):
        self.model = model.to(device)
        self.source_loader = source_loader
        self.target_loader = target_loader
        self.test_loader = test_loader
        self.num_epochs = num_epochs
        self.device = device
        self.gamma = gamma
        self.optimizer = torch.optim.Adam(model.parameters(), lr=lr)
        self.class_criterion = nn.CrossEntropyLoss()
        self.domain_criterion = nn.BCELoss()

    def _adaptive_lambda(self, epoch):
        # sigmoid 退火：训练越后期，对抗权重越接近 1
        p = epoch / self.num_epochs
        return 2.0 / (1.0 + np.exp(-self.gamma * p)) - 1.0

    def train_epoch(self, epoch):
        self.model.train()
        source_iter = iter(self.source_loader)
        target_iter = iter(self.target_loader)
        n_batches = min(len(self.source_loader), len(self.target_loader))
        total_loss = 0.0
        lambda_p = self._adaptive_lambda(epoch)

        for _ in range(n_batches):
            try:
                src_x, src_y = next(source_iter)
            except StopIteration:
                source_iter = iter(self.source_loader)
                src_x, src_y = next(source_iter)
            try:
                tgt_x, _ = next(target_iter)
            except StopIteration:
                target_iter = iter(self.target_loader)
                tgt_x, _ = next(target_iter)

            src_x = src_x.to(self.device)
            src_y = src_y.to(self.device)
            tgt_x = tgt_x.to(self.device)

            # 前向：两个头、两个域
            src_class_logits, src_dom_logits = self.model(src_x, lambda_p)
            _, tgt_dom_logits = self.model(tgt_x, lambda_p)

            # 源域分类损失
            class_loss = self.class_criterion(src_class_logits, src_y)
            # 域判别损失（源 = 1，目标 = 0）
            d_loss_s = self.domain_criterion(
                src_dom_logits, torch.ones_like(src_dom_logits))
            d_loss_t = self.domain_criterion(
                tgt_dom_logits, torch.zeros_like(tgt_dom_logits))
            domain_loss = d_loss_s + d_loss_t

            loss = class_loss + domain_loss
            self.optimizer.zero_grad()
            loss.backward()
            self.optimizer.step()
            total_loss += loss.item()

        return total_loss / n_batches

    @torch.no_grad()
    def evaluate(self):
        self.model.eval()
        preds, labels = [], []
        for x, y in self.test_loader:
            x = x.to(self.device)
            logits, _ = self.model(x, alpha=0.0)
            preds.extend(logits.argmax(dim=1).cpu().numpy())
            labels.extend(y.numpy())
        return accuracy_score(labels, preds)

    def train(self):
        best = 0.0
        for epoch in range(self.num_epochs):
            loss = self.train_epoch(epoch)
            acc = self.evaluate()
            if (epoch + 1) % 10 == 0:
                lam = self._adaptive_lambda(epoch)
                print(f"epoch {epoch + 1:3d}  loss={loss:.4f}  "
                      f"target_acc={acc:.4f}  lambda={lam:.3f}")
            best = max(best, acc)
        print(f"best target accuracy: {best:.4f}")


def main():
    N, D, C = 10000, 28 * 28, 10
    # 模拟源、目标分布偏移
    src_x = torch.randn(N, 1, 28, 28)
    src_y = torch.randint(0, C, (N,))
    tgt_x = torch.randn(N, 1, 28, 28) + 0.5     # 偏移
    tgt_y = torch.randint(0, C, (N,))           # 训练时不用
    test_x = torch.randn(2000, 1, 28, 28) + 0.5
    test_y = torch.randint(0, C, (2000,))

    BS = 128
    src_loader = DataLoader(TensorDataset(src_x, src_y), BS, shuffle=True)
    tgt_loader = DataLoader(TensorDataset(tgt_x, tgt_y), BS, shuffle=True)
    test_loader = DataLoader(TensorDataset(test_x, test_y), BS)

    model = DANN(D, 256, C)
    trainer = DANNTrainer(model, src_loader, tgt_loader, test_loader,
                          num_epochs=100, lr=1e-3)
    trainer.train()


if __name__ == "__main__":
    main()