07_finetuning.py - Examples

  1"""
  207. 파인튜닝 예제
  3
  4HuggingFace Trainer를 사용한 모델 파인튜닝
  5"""
  6
  7print("=" * 60)
  8print("파인튜닝")
  9print("=" * 60)
 10
 11
 12# ============================================
 13# 1. 기본 파인튜닝 (코드 예시)
 14# ============================================
 15print("\n[1] 기본 파인튜닝")
 16print("-" * 40)
 17
 18basic_finetuning = '''
 19from transformers import (
 20    AutoTokenizer,
 21    AutoModelForSequenceClassification,
 22    TrainingArguments,
 23    Trainer
 24)
 25from datasets import load_dataset
 26
 27# 데이터 로드
 28dataset = load_dataset("imdb")
 29
 30# 토크나이저
 31tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
 32
 33def tokenize(batch):
 34    return tokenizer(batch["text"], truncation=True, padding="max_length", max_length=256)
 35
 36tokenized = dataset.map(tokenize, batched=True)
 37
 38# 모델
 39model = AutoModelForSequenceClassification.from_pretrained("bert-base-uncased", num_labels=2)
 40
 41# 학습 설정
 42args = TrainingArguments(
 43    output_dir="./results",
 44    num_train_epochs=3,
 45    per_device_train_batch_size=16,
 46    learning_rate=2e-5,
 47    evaluation_strategy="epoch",
 48)
 49
 50# Trainer
 51trainer = Trainer(
 52    model=model,
 53    args=args,
 54    train_dataset=tokenized["train"],
 55    eval_dataset=tokenized["test"],
 56)
 57
 58# 학습
 59trainer.train()
 60'''
 61print(basic_finetuning)
 62
 63
 64# ============================================
 65# 2. LoRA 파인튜닝
 66# ============================================
 67print("\n[2] LoRA 파인튜닝")
 68print("-" * 40)
 69
 70lora_code = '''
 71from peft import LoraConfig, get_peft_model, TaskType
 72
 73# LoRA 설정
 74lora_config = LoraConfig(
 75    r=8,                           # 랭크
 76    lora_alpha=32,                 # 스케일링
 77    target_modules=["query", "value"],  # 적용 모듈
 78    lora_dropout=0.1,
 79    bias="none",
 80    task_type=TaskType.SEQ_CLS
 81)
 82
 83# 모델에 LoRA 적용
 84model = AutoModelForSequenceClassification.from_pretrained("bert-base-uncased", num_labels=2)
 85model = get_peft_model(model, lora_config)
 86
 87# 학습 가능한 파라미터 확인
 88model.print_trainable_parameters()
 89# trainable: 0.27% (약 300K / 110M)
 90
 91# 일반 Trainer로 학습
 92trainer = Trainer(model=model, args=args, ...)
 93trainer.train()
 94'''
 95print(lora_code)
 96
 97
 98# ============================================
 99# 3. QLoRA (양자화 + LoRA)
100# ============================================
101print("\n[3] QLoRA")
102print("-" * 40)
103
104qlora_code = '''
105from transformers import BitsAndBytesConfig
106import torch
107
108# 4비트 양자화 설정
109bnb_config = BitsAndBytesConfig(
110    load_in_4bit=True,
111    bnb_4bit_use_double_quant=True,
112    bnb_4bit_quant_type="nf4",
113    bnb_4bit_compute_dtype=torch.bfloat16
114)
115
116# 양자화된 모델 로드
117model = AutoModelForCausalLM.from_pretrained(
118    "meta-llama/Llama-2-7b-hf",
119    quantization_config=bnb_config,
120    device_map="auto"
121)
122
123# LoRA 적용
124model = get_peft_model(model, lora_config)
125
126# 학습
127trainer = Trainer(model=model, ...)
128'''
129print(qlora_code)
130
131
132# ============================================
133# 4. 커스텀 메트릭
134# ============================================
135print("\n[4] 커스텀 메트릭")
136print("-" * 40)
137
138try:
139    import evaluate
140    import numpy as np
141
142    # 메트릭 로드
143    accuracy = evaluate.load("accuracy")
144    f1 = evaluate.load("f1")
145
146    def compute_metrics(eval_pred):
147        logits, labels = eval_pred
148        predictions = np.argmax(logits, axis=-1)
149        return {
150            "accuracy": accuracy.compute(predictions=predictions, references=labels)["accuracy"],
151            "f1": f1.compute(predictions=predictions, references=labels, average="weighted")["f1"]
152        }
153
154    print("커스텀 메트릭 함수 정의 완료")
155
156    # 테스트
157    mock_pred = (np.array([[0.9, 0.1], [0.2, 0.8]]), np.array([0, 1]))
158    result = compute_metrics(mock_pred)
159    print(f"테스트 결과: {result}")
160
161except ImportError:
162    print("evaluate 미설치 (pip install evaluate)")
163
164
165# ============================================
166# 5. NER 파인튜닝
167# ============================================
168print("\n[5] NER 파인튜닝")
169print("-" * 40)
170
171ner_code = '''
172from transformers import AutoModelForTokenClassification
173
174# 레이블
175label_names = ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"]
176
177# 모델
178model = AutoModelForTokenClassification.from_pretrained(
179    "bert-base-uncased",
180    num_labels=len(label_names)
181)
182
183# 토큰 정렬 (서브워드 처리)
184def tokenize_and_align_labels(examples):
185    tokenized = tokenizer(examples["tokens"], truncation=True, is_split_into_words=True)
186
187    labels = []
188    for i, label in enumerate(examples["ner_tags"]):
189        word_ids = tokenized.word_ids(batch_index=i)
190        label_ids = []
191        for word_idx in word_ids:
192            if word_idx is None:
193                label_ids.append(-100)  # 특수 토큰
194            else:
195                label_ids.append(label[word_idx])
196        labels.append(label_ids)
197
198    tokenized["labels"] = labels
199    return tokenized
200'''
201print(ner_code)
202
203
204# ============================================
205# 6. QA 파인튜닝
206# ============================================
207print("\n[6] QA 파인튜닝")
208print("-" * 40)
209
210qa_code = '''
211from transformers import AutoModelForQuestionAnswering
212
213# 모델
214model = AutoModelForQuestionAnswering.from_pretrained("bert-base-uncased")
215
216# 전처리 (시작/끝 위치 찾기)
217def prepare_train_features(examples):
218    tokenized = tokenizer(
219        examples["question"],
220        examples["context"],
221        truncation="only_second",
222        max_length=384,
223        stride=128,
224        return_overflowing_tokens=True,
225        return_offsets_mapping=True,
226        padding="max_length",
227    )
228
229    # 답변 위치를 토큰 위치로 변환
230    tokenized["start_positions"] = []
231    tokenized["end_positions"] = []
232
233    for i, offsets in enumerate(tokenized["offset_mapping"]):
234        # 답변 시작/끝 문자 위치 → 토큰 위치
235        ...
236
237    return tokenized
238'''
239print(qa_code)
240
241
242# ============================================
243# 7. 학습 최적화 팁
244# ============================================
245print("\n[7] 학습 최적화 팁")
246print("-" * 40)
247
248optimization_tips = '''
249# Gradient Checkpointing (메모리 절약)
250model.gradient_checkpointing_enable()
251
252# Mixed Precision (속도 향상)
253args = TrainingArguments(
254    ...,
255    fp16=True,  # 또는 bf16=True
256)
257
258# Gradient Accumulation (큰 배치 효과)
259args = TrainingArguments(
260    per_device_train_batch_size=4,
261    gradient_accumulation_steps=8,  # 실효 배치 = 32
262)
263
264# DeepSpeed (분산 학습)
265args = TrainingArguments(
266    ...,
267    deepspeed="ds_config.json"
268)
269
270# Learning Rate Scheduler
271args = TrainingArguments(
272    learning_rate=2e-5,
273    warmup_ratio=0.1,
274    lr_scheduler_type="cosine",
275)
276'''
277print(optimization_tips)
278
279
280# ============================================
281# 정리
282# ============================================
283print("\n" + "=" * 60)
284print("파인튜닝 정리")
285print("=" * 60)
286
287summary = """
288파인튜닝 선택 가이드:
289    - 충분한 GPU: Full Fine-tuning
290    - 제한된 메모리: LoRA / QLoRA
291    - 매우 적은 데이터: Prompt Tuning
292
293핵심 코드:
294    # Trainer
295    trainer = Trainer(model=model, args=args, train_dataset=dataset)
296    trainer.train()
297
298    # LoRA
299    from peft import LoraConfig, get_peft_model
300    config = LoraConfig(r=8, target_modules=["query", "value"])
301    model = get_peft_model(model, config)
302"""
303print(summary)