Hướng dẫn

5/13•35 phút

Đang tải...

Transfer Learning & Fine-tuning

Tận dụng pre-trained models - BERT, GPT, ResNet cho domain cụ thể

Transfer Learning & Fine-tuning

Transfer Learning cho phép tận dụng knowledge từ models đã train trên dữ liệu khổng lồ cho domain riêng, tiết kiệm data và compute.

🎯 Mục tiêu

Hiểu transfer learning concept
Fine-tune pre-trained NLP models (BERT)
Fine-tune pre-trained Vision models (ResNet)
Few-shot và Zero-shot learning

1. Transfer Learning Concept

1.1 Tại sao Transfer Learning?

Train from scratch	Transfer Learning
Cần hàng triệu samples	Chỉ cần vài trăm → vài nghìn
Train vài ngày trên GPU	Fine-tune vài giờ
Chi phí cao	Chi phí thấp
Khó đạt SOTA	Gần SOTA dễ dàng

1.2 How It Works

Pre-trained Model (ImageNet/Wikipedia)

❄️Layer 1-6: General features edges, textures, shapes, grammar

🔧Layer 7-12: Task-specific features object parts, sentiment patterns

🔄Final Layer: Classification 1000 ImageNet → 5 your classes

1.3 Strategies

Strategy	When	How
Feature Extraction	Small data, similar domain	Freeze all, train new head
Fine-tuning (top)	Medium data	Freeze early, train top layers
Fine-tuning (full)	Large data, different domain	Train all with small LR

2. NLP Transfer Learning

2.1 BERT Fine-tuning (Text Classification)

Python

1from transformers import (
2    AutoTokenizer, AutoModelForSequenceClassification,
3    TrainingArguments, Trainer
4)
5from datasets import Dataset
6import pandas as pd
7
8# Load pre-trained model
9model_name = "bert-base-multilingual-cased"
10tokenizer = AutoTokenizer.from_pretrained(model_name)
11model = AutoModelForSequenceClassification.from_pretrained(
12    model_name, num_labels=3  # 3 classes: positive, neutral, negative
13)
14
15# Prepare data
16df = pd.read_csv("reviews.csv")  # columns: text, label
17dataset = Dataset.from_pandas(df)
18
19def tokenize(batch):
20    return tokenizer(batch["text"], padding=True, truncation=True, max_length=256)
21
22dataset = dataset.map(tokenize, batched=True)
23dataset = dataset.train_test_split(test_size=0.2)
24
25# Training
26training_args = TrainingArguments(
27    output_dir="./results",
28    num_train_epochs=3,
29    per_device_train_batch_size=16,
30    per_device_eval_batch_size=32,
31    learning_rate=2e-5,         # Small LR cho fine-tuning!
32    weight_decay=0.01,
33    evaluation_strategy="epoch",
34    save_strategy="epoch",
35    load_best_model_at_end=True,
36)
37
38trainer = Trainer(
39    model=model,
40    args=training_args,
41    train_dataset=dataset["train"],
42    eval_dataset=dataset["test"],
43)
44
45trainer.train()

2.2 Vietnamese NLP with PhoBERT

Python

1from transformers import AutoModel, AutoTokenizer
2
3# PhoBERT - Pre-trained on Vietnamese text
4model_name = "vinai/phobert-base-v2"
5tokenizer = AutoTokenizer.from_pretrained(model_name)
6model = AutoModel.from_pretrained(model_name)
7
8# Tokenize Vietnamese text
9text = "Sản phẩm này rất tốt, giao hàng nhanh"
10tokens = tokenizer(text, return_tensors="pt")
11outputs = model(**tokens)
12# outputs.last_hidden_state → embeddings for downstream tasks

3. Vision Transfer Learning

3.1 ResNet Fine-tuning (Image Classification)

Python

1import torch
2import torch.nn as nn
3from torchvision import models, transforms
4from torch.utils.data import DataLoader
5
6# Load pre-trained ResNet50
7model = models.resnet50(weights=models.ResNet50_Weights.IMAGENET1K_V2)
8
9# Freeze early layers
10for param in model.parameters():
11    param.requires_grad = False
12
13# Replace final layer
14num_classes = 5  # Your classes
15model.fc = nn.Sequential(
16    nn.Linear(model.fc.in_features, 256),
17    nn.ReLU(),
18    nn.Dropout(0.3),
19    nn.Linear(256, num_classes)
20)
21
22# Only train new layers
23optimizer = torch.optim.Adam(model.fc.parameters(), lr=1e-3)
24criterion = nn.CrossEntropyLoss()

3.2 Data Augmentation for Fine-tuning

Python

1train_transform = transforms.Compose([
2    transforms.RandomResizedCrop(224),
3    transforms.RandomHorizontalFlip(),
4    transforms.RandomRotation(15),
5    transforms.ColorJitter(brightness=0.2, contrast=0.2),
6    transforms.ToTensor(),
7    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
8])
9
10val_transform = transforms.Compose([
11    transforms.Resize(256),
12    transforms.CenterCrop(224),
13    transforms.ToTensor(),
14    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
15])

3.3 Progressive Unfreezing

Python

1# Phase 1: Train head only (5 epochs)
2for param in model.parameters():
3    param.requires_grad = False
4for param in model.fc.parameters():
5    param.requires_grad = True
6# Train with lr=1e-3
7
8# Phase 2: Unfreeze last block (5 epochs)
9for param in model.layer4.parameters():
10    param.requires_grad = True
11# Train with lr=1e-4 (smaller!)
12
13# Phase 3: Unfreeze all (3 epochs)
14for param in model.parameters():
15    param.requires_grad = True
16# Train with lr=1e-5 (very small!)

4. Few-shot & Zero-shot Learning

4.1 Few-shot Learning

Classify với chỉ 5-10 samples per class:

Python

1from sentence_transformers import SentenceTransformer
2from sklearn.neighbors import KNeighborsClassifier
3
4# Sentence embeddings
5model = SentenceTransformer('all-MiniLM-L6-v2')
6
7# Few examples per class
8examples = [
9    ("Great product!", "positive"),
10    ("Love it!", "positive"),
11    ("Terrible quality", "negative"),
12    ("Waste of money", "negative"),
13    ("It's okay", "neutral"),
14]
15
16texts, labels = zip(*examples)
17embeddings = model.encode(list(texts))
18
19# KNN classifier on embeddings
20knn = KNeighborsClassifier(n_neighbors=3)
21knn.fit(embeddings, labels)
22
23# Classify new text
24new_text = "Absolutely amazing experience"
25new_embed = model.encode([new_text])
26print(knn.predict(new_embed))  # "positive"

4.2 Zero-shot Classification

Python

1from transformers import pipeline
2
3classifier = pipeline("zero-shot-classification",
4                      model="facebook/bart-large-mnli")
5
6text = "Chiếc điện thoại này pin rất trâu"
7labels = ["tích cực", "tiêu cực", "trung lập"]
8
9result = classifier(text, candidate_labels=labels)
10print(result)
11# {'labels': ['tích cực', 'trung lập', 'tiêu cực'],
12#  'scores': [0.85, 0.12, 0.03]}

5. Best Practices

5.1 Fine-tuning Checklist

Step	Detail
1. Choose base model	Match domain (NLP/Vision/Tabular)
2. Prepare data	Clean, labeled, balanced
3. Learning rate	Start small (2e-5 for BERT, 1e-3 for head)
4. Freeze strategy	Start frozen, progressively unfreeze
5. Regularization	Dropout, weight decay
6. Evaluation	Hold-out + cross-validation
7. Early stopping	Monitor val loss, patience=3

5.2 Common LR Schedule

Python

1from transformers import get_linear_schedule_with_warmup
2
3scheduler = get_linear_schedule_with_warmup(
4    optimizer,
5    num_warmup_steps=100,     # Warm up slowly
6    num_training_steps=1000   # Then decay
7)

📝 Quiz

🎯 Key Takeaways

Transfer Learning tiết kiệm data + compute đáng kể
Freeze → Unfreeze progressively cho best results
Small LR critical cho fine-tuning
PhoBERT — best choice cho Vietnamese NLP
Few-shot/Zero-shot — powerful khi data rất ít

🚀 Bài tiếp theo

Recommendation Systems Overview — Hệ thống gợi ý: Collaborative Filtering, Content-based, và Hybrid!

AI Trợ lý

Transfer Learning & Fine-tuning

Transfer Learning & Fine-tuning

🎯 Mục tiêu

1. Transfer Learning Concept

1.1 Tại sao Transfer Learning?

1.2 How It Works

Pre-trained Model (ImageNet/Wikipedia)

1.3 Strategies

2. NLP Transfer Learning

2.1 BERT Fine-tuning (Text Classification)

2.2 Vietnamese NLP with PhoBERT

3. Vision Transfer Learning

3.1 ResNet Fine-tuning (Image Classification)

3.2 Data Augmentation for Fine-tuning

3.3 Progressive Unfreezing

4. Few-shot & Zero-shot Learning

4.1 Few-shot Learning

4.2 Zero-shot Classification

5. Best Practices

5.1 Fine-tuning Checklist

5.2 Common LR Schedule

📝 Quiz

🎯 Key Takeaways

🚀 Bài tiếp theo

Khóa học

Mentor & Hỗ trợ

Blog

Giới thiệu