迁移学习（二）：预训练与微调

  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

import torch
import torch.nn as nn
from torch.utils.data import DataLoader, Dataset
from transformers import (
    BertTokenizer, BertModel, AdamW, get_linear_schedule_with_warmup,
)
from torch.cuda.amp import autocast, GradScaler
from tqdm import tqdm
from sklearn.metrics import accuracy_score, f1_score

class BERTClassifier(nn.Module):
    """BERT 主干网络 + [CLS] 上的线性分类头。"""

    def __init__(self, bert_model_name="bert-base-uncased",
                 num_classes=2, dropout=0.1):
        super().__init__()
        self.bert = BertModel.from_pretrained(bert_model_name)
        self.dropout = nn.Dropout(dropout)
        self.classifier = nn.Linear(self.bert.config.hidden_size, num_classes)

    def forward(self, input_ids, attention_mask):
        outputs = self.bert(input_ids=input_ids, attention_mask=attention_mask)
        pooled = self.dropout(outputs.pooler_output)
        return self.classifier(pooled)

class TextDataset(Dataset):
    """对文本进行分词和填充，适配 BERT 输入格式。"""

    def __init__(self, texts, labels, tokenizer, max_length=128):
        self.texts = texts
        self.labels = labels
        self.tokenizer = tokenizer
        self.max_length = max_length

    def __len__(self):
        return len(self.texts)

    def __getitem__(self, idx):
        encoding = self.tokenizer.encode_plus(
            str(self.texts[idx]),
            add_special_tokens=True,
            max_length=self.max_length,
            padding="max_length",
            truncation=True,
            return_attention_mask=True,
            return_tensors="pt",
        )
        return {
            "input_ids": encoding["input_ids"].flatten(),
            "attention_mask": encoding["attention_mask"].flatten(),
            "label": torch.tensor(self.labels[idx], dtype=torch.long),
        }

class BERTFineTuner:
    """支持判别式学习率、梯度累积和混合精度的训练器。"""

    def __init__(self, model, train_loader, val_loader,
                 num_epochs=3, learning_rate=2e-5, warmup_ratio=0.1,
                 gradient_accumulation_steps=1, max_grad_norm=1.0,
                 device="cuda", use_amp=True,
                 discriminative_lr=False, lr_decay=2.6):
        self.model = model.to(device)
        self.train_loader = train_loader
        self.val_loader = val_loader
        self.num_epochs = num_epochs
        self.device = device
        self.use_amp = use_amp
        self.gradient_accumulation_steps = gradient_accumulation_steps
        self.max_grad_norm = max_grad_norm

        total_steps = len(train_loader) * num_epochs // gradient_accumulation_steps
        warmup_steps = int(total_steps * warmup_ratio)

        if discriminative_lr:
            self.optimizer = self._discriminative_optimizer(learning_rate, lr_decay)
        else:
            self.optimizer = AdamW(model.parameters(), lr=learning_rate, eps=1e-8)

        self.scheduler = get_linear_schedule_with_warmup(
            self.optimizer, warmup_steps, total_steps,
        )
        self.scaler = GradScaler() if use_amp else None
        self.criterion = nn.CrossEntropyLoss()

    def _discriminative_optimizer(self, lr, decay):
        """底层学习率指数级递减，顶层保持最大学习率。"""
        num_layers = len(self.model.bert.encoder.layer)
        groups = []
        # Embedding 层 —— 学习率最小
        groups.append({
            "params": self.model.bert.embeddings.parameters(),
            "lr": lr / (decay ** num_layers),
        })
        # 每个 Transformer 层
        for i in range(num_layers):
            groups.append({
                "params": self.model.bert.encoder.layer[i].parameters(),
                "lr": lr / (decay ** (num_layers - i - 1)),
            })
        # Pooler + 分类头 —— 学习率最大
        groups.append({
            "params": list(self.model.bert.pooler.parameters())
                      + list(self.model.classifier.parameters()),
            "lr": lr,
        })
        return AdamW(groups, eps=1e-8)

    def train_epoch(self):
        self.model.train()
        total_loss = 0.0
        for step, batch in enumerate(tqdm(self.train_loader, desc="训练中")):
            input_ids = batch["input_ids"].to(self.device)
            mask = batch["attention_mask"].to(self.device)
            labels = batch["label"].to(self.device)

            if self.use_amp:
                with autocast():
                    loss = self.criterion(self.model(input_ids, mask), labels)
                loss = loss / self.gradient_accumulation_steps
                self.scaler.scale(loss).backward()
            else:
                loss = self.criterion(self.model(input_ids, mask), labels)
                loss = loss / self.gradient_accumulation_steps
                loss.backward()

            if (step + 1) % self.gradient_accumulation_steps == 0:
                if self.use_amp:
                    self.scaler.unscale_(self.optimizer)
                    nn.utils.clip_grad_norm_(self.model.parameters(),
                                             self.max_grad_norm)
                    self.scaler.step(self.optimizer)
                    self.scaler.update()
                else:
                    nn.utils.clip_grad_norm_(self.model.parameters(),
                                             self.max_grad_norm)
                    self.optimizer.step()
                self.scheduler.step()
                self.optimizer.zero_grad()

            total_loss += loss.item() * self.gradient_accumulation_steps
        return total_loss / len(self.train_loader)

    @torch.no_grad()
    def evaluate(self):
        self.model.eval()
        all_preds, all_labels, total_loss = [], [], 0.0
        for batch in tqdm(self.val_loader, desc="评估中"):
            input_ids = batch["input_ids"].to(self.device)
            mask = batch["attention_mask"].to(self.device)
            labels = batch["label"].to(self.device)
            logits = self.model(input_ids, mask)
            total_loss += self.criterion(logits, labels).item()
            all_preds.extend(logits.argmax(dim=1).cpu().numpy())
            all_labels.extend(labels.cpu().numpy())
        return (
            total_loss / len(self.val_loader),
            accuracy_score(all_labels, all_preds),
            f1_score(all_labels, all_preds, average="weighted"),
        )

    def train(self):
        best_val_loss = float("inf")
        for epoch in range(self.num_epochs):
            train_loss = self.train_epoch()
            val_loss, val_acc, val_f1 = self.evaluate()
            print(f"Epoch {epoch + 1}: train_loss={train_loss:.4f}  "
                  f"val_loss={val_loss:.4f}  acc={val_acc:.4f}  f1={val_f1:.4f}")
            if val_loss < best_val_loss:
                best_val_loss = val_loss
                torch.save(self.model.state_dict(), "best_model.pt")
                print("  -> 已保存最佳模型")

def main():
    BERT_MODEL = "bert-base-uncased"
    tokenizer = BertTokenizer.from_pretrained(BERT_MODEL)

    # 假数据，请替换为真实数据集
    texts = ["This movie is great!"] * 500 + ["This movie is terrible!"] * 500
    labels = [1] * 500 + [0] * 500

    dataset = TextDataset(texts, labels, tokenizer, max_length=128)
    train_loader = DataLoader(dataset, batch_size=16, shuffle=True)
    val_loader = DataLoader(dataset, batch_size=16)

    model = BERTClassifier(BERT_MODEL, num_classes=2)
    trainer = BERTFineTuner(
        model, train_loader, val_loader,
        num_epochs=3, learning_rate=2e-5,
        gradient_accumulation_steps=2,
        discriminative_lr=True,
        use_amp=True,
    )
    trainer.train()

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
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56

import torch
import torch.nn.functional as F
from torch.utils.data import DataLoader

def compute_fisher(model, dataloader, n_samples=500, device="cuda"):
    """在当前参数处计算对角 Fisher 信息。"""
    model.eval()
    fisher = {n: torch.zeros_like(p) for n, p in model.named_parameters()
              if p.requires_grad}
    seen = 0
    for x, y in dataloader:
        x, y = x.to(device), y.to(device)
        model.zero_grad()
        logits = model(x)
        # 从模型预测分布中采样 y_hat。
        log_probs = F.log_softmax(logits, dim=-1)
        y_hat = torch.multinomial(log_probs.exp(), 1).squeeze(-1)
        loss = F.nll_loss(log_probs, y_hat, reduction="sum")
        loss.backward()
        for n, p in model.named_parameters():
            if p.grad is not None:
                fisher[n] += p.grad.detach() ** 2
        seen += x.size(0)
        if seen >= n_samples:
            break
    for n in fisher:
        fisher[n] /= max(seen, 1)
    return fisher

def ewc_penalty(model, theta_star, fisher, lam=400.0):
    """使用对角 Fisher 加权的二次锚定惩罚。"""
    loss = 0.0
    for n, p in model.named_parameters():
        if n in fisher:
            loss = loss + (fisher[n] * (p - theta_star[n]) ** 2).sum()
    return 0.5 * lam * loss

def finetune_with_ewc(model, source_loader, target_loader,
                      epochs=10, lr=1e-3, lam=400.0, device="cuda"):
    """两阶段：保存 theta*，在任务 1 上计算 Fisher，然后在任务 2 上微调。"""
    theta_star = {n: p.detach().clone()
                  for n, p in model.named_parameters() if p.requires_grad}
    fisher = compute_fisher(model, source_loader, n_samples=500, device=device)

    opt = torch.optim.SGD(model.parameters(), lr=lr, momentum=0.9)
    for epoch in range(epochs):
        model.train()
        for x, y in target_loader:
            x, y = x.to(device), y.to(device)
            logits = model(x)
            loss = F.cross_entropy(logits, y) \
                 + ewc_penalty(model, theta_star, fisher, lam=lam)
            opt.zero_grad()
            loss.backward()
            opt.step()
    return model

 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

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

def nt_xent(z1, z2, tau):
    """带可配置温度的 SimCLR 损失。"""
    z = torch.cat([z1, z2], dim=0)                  # 2B x d
    z = F.normalize(z, dim=-1)
    sim = z @ z.t() / tau                           # 2B x 2B
    B = z1.size(0)
    mask = torch.eye(2 * B, device=z.device, dtype=torch.bool)
    sim.masked_fill_(mask, -1e9)                    # 去除自相似
    targets = torch.cat([torch.arange(B, 2 * B),
                         torch.arange(0, B)]).to(z.device)
    return F.cross_entropy(sim, targets)

def tau_schedule(step, total_steps, tau_start=0.5, tau_end=0.07, warmup_frac=0.1):
    """温度余弦预热：初始较大，快速退火。"""
    warmup_steps = int(total_steps * warmup_frac)
    if step >= warmup_steps:
        return tau_end
    progress = step / max(warmup_steps, 1)
    return tau_end + 0.5 * (tau_start - tau_end) * (1 + math.cos(math.pi * progress))

def train_step(model, x1, x2, opt, step, total_steps):
    tau = tau_schedule(step, total_steps)
    z1, z2 = model(x1), model(x2)
    loss = nt_xent(z1, z2, tau)
    opt.zero_grad()
    loss.backward()
    grad_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1e9)
    opt.step()
    if step % 50 == 0:
        print(f"step {step}  tau={tau:.3f}  loss={loss.item():.3f}  "
              f"grad_norm={grad_norm.item():.2f}")
    return loss.item(), grad_norm.item()

 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

import torch
import torch.nn as nn

class LoRALinear(nn.Module):
    """冻结的 Linear 层，在前向时叠加一个或多个 LoRA 增量。"""
    def __init__(self, base: nn.Linear):
        super().__init__()
        self.base = base
        for p in self.base.parameters():
            p.requires_grad = False
        self.adapters = nn.ModuleList()  # 每个为 (A, B) 对
        self.weights = []                # 每个适配器的 alpha

    def add_adapter(self, A: torch.Tensor, B: torch.Tensor, alpha: float):
        ad = nn.ParameterDict({
            "A": nn.Parameter(A, requires_grad=False),
            "B": nn.Parameter(B, requires_grad=False),
        })
        self.adapters.append(ad)
        self.weights.append(alpha)

    def forward(self, x):
        out = self.base(x)
        for ad, alpha in zip(self.adapters, self.weights):
            out = out + alpha * (x @ ad["A"].t() @ ad["B"].t())
        return out

def compose_loras(base_model, lora_list, weights):
    """将 lora_list 中指定的 Linear 层替换为 LoRALinear，
    并按给定权重叠加提供的 (A, B) 增量。"""
    name_to_loras = {}  # layer_name -> (A, B) 列表
    for lora in lora_list:
        for name, (A, B) in lora.items():
            name_to_loras.setdefault(name, []).append((A, B))

    for name, mod in list(base_model.named_modules()):
        if name in name_to_loras and isinstance(mod, nn.Linear):
            wrapped = LoRALinear(mod)
            for (A, B), alpha in zip(name_to_loras[name], weights):
                wrapped.add_adapter(A, B, alpha)
            parent = base_model
            *path, last = name.split(".")
            for p in path:
                parent = getattr(parent, p)
            setattr(parent, last, wrapped)
    return base_model