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
197
198
199
200
201

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 骨干 + [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):
    """对文本做分词与 padding，喂给 BERT。"""

    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:
    """带判别式 LR、梯度累积、AMP 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):
        """给底层指数级地减小学习率。"""
        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 + 分类器 —— 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="Training")):
            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="Evaluating"):
            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("  -> 已保存当前最佳模型")


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

    # 假数据，请替换成真实数据集
    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()