Machine Learning Academy · 课时

微调 BertForSequenceClassification

您将加载预训练的 BERT 检查点，添加分类头，创建 PyTorch DataLoader，并在 IMDB 数据集上微调两个训练轮次。

微调 BertForSequenceClassification 是 CoddyKit 上的免费 Machine Learning Academy 课时。这是第 3 节课，共 4 节。你可以在下方免费阅读本课时的完整内容 — 然后在浏览器中使用内置代码编辑器和全天候 AI 导师进行实践。这是 Machine Learning Academy 学习路径的一部分，你的进度在网页和 CoddyKit 应用中同步。 Machine Learning Academy 课程共包含 4 节课。

本课时的部分内容尚未翻译，以英文显示。

What Is Fine-Tuning?

Fine-tuning takes a pre-trained model that already understands language structure and adapts it to a specific task with a small labelled dataset. BERT pre-trained on 3.3 billion words already knows grammar, semantics, and world knowledge. Fine-tuning adds a task-specific head (e.g., a classification layer) and trains the entire model end-to-end on your labelled data for a few epochs, achieving state-of-the-art results with far less data and compute than training from scratch.

BertForSequenceClassification Overview

BertForSequenceClassification is a BERT model with a linear classification head on top of the [CLS] token's final hidden state. It is the standard Hugging Face class for sentiment analysis, topic classification, and any task that assigns a single label to an entire text. The head is initialised randomly and trained alongside the BERT backbone during fine-tuning.

from transformers import BertForSequenceClassification
import torch

# 2 classes: negative (0) and positive (1)
model = BertForSequenceClassification.from_pretrained(
    'bert-base-uncased',
    num_labels=2
)
print(model.config.num_labels)   # 2
print(model.classifier)          # Linear(in=768, out=2)

Loading the IMDB Dataset

The IMDB dataset contains 50,000 movie reviews labelled positive (1) or negative (0). We use the Hugging Face datasets library to download and split it. The dataset object behaves like a dictionary of lists and can be mapped over with a tokenization function. We use a 25k train / 25k test split provided by the original dataset authors.

from datasets import load_dataset
from transformers import BertTokenizer

dataset = load_dataset('imdb')
print(dataset)  # DatasetDict with train/test splits

tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')

def tokenize_fn(examples):
    return tokenizer(
        examples['text'],
        truncation=True,
        padding='max_length',
        max_length=256
    )

tokenized = dataset.map(tokenize_fn, batched=True)
tokenized.set_format('torch', columns=['input_ids', 'attention_mask', 'label'])

Creating PyTorch DataLoaders

After tokenisation, wrap the Hugging Face dataset in PyTorch DataLoader objects. The batch_size controls how many examples the model processes per forward pass; 16 or 32 is typical for BERT given GPU memory constraints. Shuffle the training set each epoch so the model does not overfit to example order, but keep the validation set fixed for reproducible evaluation.

from torch.utils.data import DataLoader

train_dataset = tokenized['train'].select(range(2000))  # small subset for demo
test_dataset = tokenized['test'].select(range(500))

train_loader = DataLoader(train_dataset, batch_size=16, shuffle=True)
test_loader = DataLoader(test_dataset, batch_size=32, shuffle=False)

print('Train batches:', len(train_loader))
print('Test batches:', len(test_loader))

Setting Up the Optimizer

Fine-tuning BERT uses the AdamW optimizer with a small learning rate (typically 2e-5 to 5e-5). Using a large learning rate destroys the pre-trained representations (catastrophic forgetting). Weight decay (L2 regularisation) is applied to non-bias parameters to reduce overfitting. Hugging Face's get_linear_schedule_with_warmup is commonly used to warm up the LR over the first 10% of steps then decay linearly.

from torch.optim import AdamW
from transformers import get_linear_schedule_with_warmup

optimizer = AdamW(model.parameters(), lr=2e-5, weight_decay=0.01)

num_epochs = 2
num_steps = len(train_loader) * num_epochs
warmup_steps = int(0.1 * num_steps)

scheduler = get_linear_schedule_with_warmup(
    optimizer,
    num_warmup_steps=warmup_steps,
    num_training_steps=num_steps
)

The Fine-Tuning Training Loop

The PyTorch training loop for fine-tuning BERT follows the standard pattern: iterate over batches, compute the forward pass, extract the loss from the model output (BertForSequenceClassification returns loss automatically when labels are passed), call loss.backward(), clip gradients to prevent explosion, and step the optimizer and scheduler. Training BERT for 2 epochs on IMDB typically yields ~92% accuracy.

import torch

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = model.to(device)

for epoch in range(2):
    model.train()
    total_loss = 0
    for batch in train_loader:
        optimizer.zero_grad()
        input_ids = batch['input_ids'].to(device)
        attention_mask = batch['attention_mask'].to(device)
        labels = batch['label'].to(device)

        outputs = model(input_ids=input_ids,
                        attention_mask=attention_mask,
                        labels=labels)
        loss = outputs.loss
        loss.backward()
        torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
        optimizer.step()
        scheduler.step()
        total_loss += loss.item()

    print(f'Epoch {epoch+1} loss: {total_loss/len(train_loader):.4f}')

Evaluating on the Validation Set

During evaluation, call model.eval() to disable dropout and use torch.no_grad() to skip gradient computation, saving memory and speeding up inference. Extract logits from the model output (raw scores before softmax), apply argmax to get predicted class indices, and compare against ground-truth labels to compute accuracy.

import torch
from sklearn.metrics import accuracy_score

model.eval()
all_preds, all_labels = [], []

with torch.no_grad():
    for batch in test_loader:
        input_ids = batch['input_ids'].to(device)
        attention_mask = batch['attention_mask'].to(device)
        labels = batch['label'].to(device)

        outputs = model(input_ids=input_ids,
                        attention_mask=attention_mask)
        preds = torch.argmax(outputs.logits, dim=1)
        all_preds.extend(preds.cpu().numpy())
        all_labels.extend(labels.cpu().numpy())

print('Accuracy:', accuracy_score(all_labels, all_preds))

Saving the Fine-Tuned Model

After fine-tuning, save the model and tokenizer together so they can be reloaded for inference without re-training. model.save_pretrained(path) saves the model weights and config; tokenizer.save_pretrained(path) saves the vocabulary and tokenization settings. Reload with from_pretrained(path) in any new Python session.

import os

save_dir = './bert_imdb_finetuned'
os.makedirs(save_dir, exist_ok=True)

model.save_pretrained(save_dir)
tokenizer.save_pretrained(save_dir)
print('Model saved to', save_dir)

# Reload in a new session:
# from transformers import BertForSequenceClassification, BertTokenizer
# model = BertForSequenceClassification.from_pretrained(save_dir)
# tokenizer = BertTokenizer.from_pretrained(save_dir)

Using the Trainer API

Hugging Face's Trainer class encapsulates the training loop, evaluation, checkpointing, and logging in a single high-level API. Define a TrainingArguments object with learning rate, batch size, and number of epochs, then call trainer.train(). The Trainer handles gradient clipping, LR scheduling, and mixed-precision training automatically, reducing boilerplate significantly.

from transformers import Trainer, TrainingArguments
import numpy as np
from sklearn.metrics import accuracy_score

def compute_metrics(eval_pred):
    logits, labels = eval_pred
    preds = np.argmax(logits, axis=1)
    return {'accuracy': accuracy_score(labels, preds)}

training_args = TrainingArguments(
    output_dir='./bert_trainer',
    num_train_epochs=2,
    per_device_train_batch_size=16,
    per_device_eval_batch_size=32,
    learning_rate=2e-5,
    evaluation_strategy='epoch',
    save_strategy='epoch',
    load_best_model_at_end=True
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=test_dataset,
    compute_metrics=compute_metrics
)
trainer.train()

Monitoring Training with TensorBoard

Pass report_to='tensorboard' in TrainingArguments to log loss and metrics to TensorBoard automatically. Launch TensorBoard with tensorboard --logdir ./bert_trainer/runs in a terminal to view real-time loss curves, learning rate schedules, and evaluation metrics. Monitoring training loss vs validation loss helps you detect overfitting early and stop training at the right epoch.

from transformers import TrainingArguments

training_args = TrainingArguments(
    output_dir='./bert_trainer',
    num_train_epochs=3,
    per_device_train_batch_size=16,
    learning_rate=2e-5,
    evaluation_strategy='steps',
    eval_steps=100,
    logging_steps=50,
    report_to='tensorboard',   # enable TensorBoard logging
    logging_dir='./bert_trainer/runs'
)
# Then: tensorboard --logdir ./bert_trainer/runs

Choosing How Many Layers to Freeze

Sometimes the dataset is too small to fine-tune all 12 BERT layers without overfitting. A common strategy is to freeze the lower layers (which learn general linguistic features) and only update the upper layers (which learn task-specific features). Freeze by setting requires_grad=False on selected parameter groups. Monitor validation accuracy to choose the optimal freezing depth for your dataset size.

from transformers import BertForSequenceClassification

model = BertForSequenceClassification.from_pretrained('bert-base-uncased', num_labels=2)

# Freeze embedding layer and first 6 encoder layers
for name, param in model.named_parameters():
    if 'embeddings' in name or 'encoder.layer.0' in name or \
       'encoder.layer.1' in name or 'encoder.layer.2' in name:
        param.requires_grad = False

# Count trainable parameters
trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
print(f'Trainable parameters: {trainable:,}')

Quick Check

Test your understanding of Machine Learning with Python concepts from this lesson.

Lesson Recap

In this lesson you learned: BertForSequenceClassification adds a linear head on [CLS] for classification tasks, fine-tuning uses AdamW with a very small learning rate to avoid catastrophic forgetting, and the Hugging Face Trainer API simplifies the training loop with built-in evaluation and checkpointing. Next up we look at how to extract predictions from logits and evaluate the fine-tuned model with accuracy and F1.

常见问题解答

「微调 BertForSequenceClassification」课时是免费的吗？

是的 — 「微调 BertForSequenceClassification」的完整文本可在网页上免费阅读。要进行交互式练习（内置代码编辑器和全天候 AI 导师）并解锁 Machine Learning Academy 课程的其余内容，请升级到 CoddyKit PRO。 Machine Learning Academy 课程共包含 4 节课。

「微调 BertForSequenceClassification」这节课中我会学到什么？

您将加载预训练的 BERT 检查点，添加分类头，创建 PyTorch DataLoader，并在 IMDB 数据集上微调两个训练轮次。你通过在浏览器中直接运行的动手代码来练习 Machine Learning Academy，全天候 AI 导师会在你学习这节课的过程中回答你的问题。

学习 Machine Learning Academy 需要有经验吗？

无需任何先前经验。CoddyKit 上的 Machine Learning Academy 课程适合初学者到高级学习者，你可以从这里开始或从头开始，按照自己的节奏学习。这是第 3 节课，共 4 节。

「微调 BertForSequenceClassification」课时需要多长时间？

大多数 CoddyKit 课程大约需要 5–10 分钟。每节课都很精短且互动，所以你能稳步进步，并在网页和应用中从离开的地方继续。

我能在这节 Machine Learning Academy 课中编写并运行代码吗？

能。每节 Machine Learning Academy 课都包含内置代码编辑器，你可以在浏览器中直接编写并运行真实代码，并获得即时 AI 反馈 — 无需本地设置。

此课程中的所有课时

← 返回 Machine Learning Academy