huggingface_transformers（拥抱脸变形金刚）

Hugging Face Transformers Best Practices

Comprehensive guide to using the Hugging Face Transformers library including model loading, tokenization, fine-tuning workflows, pipeline usage, custom datasets, and deployment optimization.

Quick Reference

When to use this skill:

• Loading and using pre-trained transformers (BERT, GPT, T5, LLaMA, etc.)
• Fine-tuning models on custom data
• Implementing NLP tasks (classification, QA, generation, etc.)
• Optimizing inference (quantization, ONNX, etc.)
• Debugging tokenization issues
• Using Hugging Face pipelines
• Deploying transformers to production

Models covered:

• Encoders: BERT, RoBERTa, DeBERTa, ALBERT
• Decoders: GPT-2, GPT-Neo, LLaMA, Mistral
• Encoder-Decoders: T5, BART, Flan-T5
• Vision: ViT, CLIP, Stable Diffusion

Part 1: Model Loading Patterns

Pattern 1: Basic Model Loading

from transformers import AutoModel, AutoTokenizer

# Load model and tokenizer
model_name = "bert-base-uncased"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModel.from_pretrained(model_name)

# For specific tasks
from transformers import AutoModelForSequenceClassification
model = AutoModelForSequenceClassification.from_pretrained(
    model_name,
    num_labels=3  # For 3-class classification
)

Pattern 2: Loading with Specific Configuration

from transformers import AutoConfig, AutoModel

# Modify configuration
config = AutoConfig.from_pretrained("bert-base-uncased")
config.hidden_dropout_prob = 0.2  # Custom dropout
config.attention_probs_dropout_prob = 0.2

# Load model with custom config
model = AutoModel.from_pretrained("bert-base-uncased", config=config)

# Or create model from scratch with config
model = AutoModel.from_config(config)

Pattern 3: Loading Quantized Models (Memory Efficient)

from transformers import AutoModel, BitsAndBytesConfig
import torch

# 8-bit quantization (50% memory reduction)
quantization_config = BitsAndBytesConfig(load_in_8bit=True)

model = AutoModel.from_pretrained(
    "meta-llama/Llama-2-7b-hf",
    quantization_config=quantization_config,
    device_map="auto"  # Automatic device placement
)

# 4-bit quantization (75% memory reduction)
quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True
)

model = AutoModel.from_pretrained(
    "meta-llama/Llama-2-13b-hf",
    quantization_config=quantization_config,
    device_map="auto"
)

Pattern 4: Loading from Local Path

# Save model locally
model.save_pretrained("./my-model")
tokenizer.save_pretrained("./my-model")

# Load from local path
model = AutoModel.from_pretrained("./my-model")
tokenizer = AutoTokenizer.from_pretrained("./my-model")

Part 2: Tokenization Best Practices

Critical Tokenization Patterns

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")

# ✅ CORRECT: All required arguments
tokens = tokenizer(
    text,
    padding=True,  # Pad to longest in batch
    truncation=True,  # Truncate to max_length
    max_length=512,  # Maximum sequence length
    return_tensors="pt"  # Return PyTorch tensors
)

# Access components
input_ids = tokens['input_ids']  # Token IDs
attention_mask = tokens['attention_mask']  # Padding mask
token_type_ids = tokens.get('token_type_ids')  # Segment IDs (BERT)

# ❌ WRONG: Missing critical arguments
tokens = tokenizer(text)  # No padding, truncation, or tensor format!

Batch Tokenization

# Tokenize multiple texts efficiently
texts = ["First text", "Second text", "Third text"]

tokens = tokenizer(
    texts,
    padding=True,  # Pad all to longest in batch
    truncation=True,
    max_length=128,
    return_tensors="pt"
)

# Result shape: [batch_size, max_length]
print(tokens['input_ids'].shape)  # torch.Size([3, max_len_in_batch])

Special Token Handling

# Add special tokens
tokenizer.add_special_tokens({
    'additional_special_tokens': ['[CUSTOM]', '[MARKER]']
})

# Resize model embeddings to match
model.resize_token_embeddings(len(tokenizer))

# Encode with special tokens preserved
text = "Hello [CUSTOM] world"
tokens = tokenizer(text, add_special_tokens=True)

# Decode
decoded = tokenizer.decode(tokens['input_ids'][0], skip_special_tokens=False)

Tokenization for Different Tasks

# Text classification (single sequence)
tokens = tokenizer(
    "This movie was great!",
    padding="max_length",
    truncation=True,
    max_length=128,
    return_tensors="pt"
)

# Question answering (pair of sequences)
question = "What is the capital of France?"
context = "France is a country in Europe. Paris is its capital."

tokens = tokenizer(
    question,
    context,
    padding="max_length",
    truncation="only_second",  # Only truncate context
    max_length=384,
    return_tensors="pt"
)

# Text generation (decoder-only models)
prompt = "Once upon a time"
tokens = tokenizer(prompt, return_tensors="pt")
# No padding needed for generation input

Part 3: Fine-Tuning Workflows

Pattern 1: Simple Fine-Tuning with Trainer

from transformers import (
    AutoModelForSequenceClassification,
    AutoTokenizer,
    Trainer,
    TrainingArguments
)
from datasets import load_dataset

# 1. Load dataset
dataset = load_dataset("glue", "mrpc")

# 2. Load model
model = AutoModelForSequenceClassification.from_pretrained(
    "bert-base-uncased",
    num_labels=2
)
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")

# 3. Tokenize dataset
def tokenize_function(examples):
    return tokenizer(
        examples["sentence1"],
        examples["sentence2"],
        padding="max_length",
        truncation=True,
        max_length=128
    )

tokenized_datasets = dataset.map(tokenize_function, batched=True)

# 4. Define training arguments
training_args = TrainingArguments(
    output_dir="./results",
    evaluation_strategy="epoch",
    learning_rate=2e-5,
    per_device_train_batch_size=16,
    per_device_eval_batch_size=16,
    num_train_epochs=3,
    weight_decay=0.01,
    logging_dir="./logs",
    logging_steps=100,
    save_strategy="epoch",
    load_best_model_at_end=True,
    metric_for_best_model="accuracy",
)

# 5. Define metrics
from datasets import load_metric
import numpy as np

metric = load_metric("accuracy")

def compute_metrics(eval_pred):
    logits, labels = eval_pred
    predictions = np.argmax(logits, axis=-1)
    return metric.compute(predictions=predictions, references=labels)

# 6. Create Trainer
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_datasets["train"],
    eval_dataset=tokenized_datasets["validation"],
    compute_metrics=compute_metrics,
)

# 7. Train
trainer.train()

# 8. Save
trainer.save_model("./fine-tuned-model")

Pattern 2: LoRA Fine-Tuning (Parameter-Efficient)

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import LoraConfig, get_peft_model, TaskType

# Load base model
model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-2-7b-hf",
    load_in_8bit=True,  # 8-bit for memory efficiency
    device_map="auto"
)

# Configure LoRA
lora_config = LoraConfig(
    task_type=TaskType.CAUSAL_LM,
    r=8,  # LoRA rank
    lora_alpha=32,  # LoRA alpha
    lora_dropout=0.1,
    target_modules=["q_proj", "v_proj"],  # Which layers to adapt
)

# Apply LoRA
model = get_peft_model(model, lora_config)

# Check trainable parameters
model.print_trainable_parameters()
# Output: trainable params: 4.2M || all params: 6.7B || trainable%: 0.062%

# Train with Trainer (same as before)
# Only LoRA parameters are updated!

Pattern 3: Custom Training Loop

import torch
from torch.utils.data import DataLoader
from transformers import AdamW, get_scheduler

# Prepare dataloaders
train_dataloader = DataLoader(tokenized_datasets["train"], batch_size=16, shuffle=True)
eval_dataloader = DataLoader(tokenized_datasets["validation"], batch_size=16)

# Optimizer
optimizer = AdamW(model.parameters(), lr=2e-5)

# Learning rate scheduler
num_epochs = 3
num_training_steps = num_epochs * len(train_dataloader)
lr_scheduler = get_scheduler(
    "linear",
    optimizer=optimizer,
    num_warmup_steps=500,
    num_training_steps=num_training_steps
)

# Training loop
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

for epoch in range(num_epochs):
    model.train()
    for batch in train_dataloader:
        batch = {k: v.to(device) for k, v in batch.items()}

        outputs = model(**batch)
        loss = outputs.loss
        loss.backward()

        optimizer.step()
        lr_scheduler.step()
        optimizer.zero_grad()

    # Evaluation
    model.eval()
    for batch in eval_dataloader:
        batch = {k: v.to(device) for k, v in batch.items()}
        with torch.no_grad():
            outputs = model(**batch)
        # Compute metrics

Part 4: Pipeline Usage (High-Level API)

Text Classification Pipeline

from transformers import pipeline

# Load pipeline
classifier = pipeline(
    "text-classification",
    model="distilbert-base-uncased-finetuned-sst-2-english"
)

# Single prediction
result = classifier("I love this product!")
# [{'label': 'POSITIVE', 'score': 0.9998}]

# Batch prediction
results = classifier([
    "Great service!",
    "Terrible experience",
    "Average quality"
])

Question Answering Pipeline

qa_pipeline = pipeline("question-answering", model="distilbert-base-uncased-distilled-squad")

result = qa_pipeline(
    question="What is the capital of France?",
    context="France is a country in Europe. Its capital is Paris, a beautiful city."
)
# {'score': 0.98, 'start': 49, 'end': 54, 'answer': 'Paris'}

Text Generation Pipeline

generator = pipeline("text-generation", model="gpt2")

outputs = generator(
    "Once upon a time",
    max_length=50,
    num_return_sequences=3,
    temperature=0.7,
    top_k=50,
    top_p=0.95,
    do_sample=True
)

for output in outputs:
    print(output['generated_text'])

Zero-Shot Classification Pipeline

classifier = pipeline("zero-shot-classification", model="facebook/bart-large-mnli")

result = classifier(
    "This is a course about Python programming.",
    candidate_labels=["education", "technology", "business", "sports"]
)
# {'sequence': '...', 'labels': ['education', 'technology', ...], 'scores': [0.85, 0.12, ...]}

Part 5: Inference Optimization

Optimization 1: Batch Processing

# ❌ SLOW: Process one at a time
for text in texts:
    output = model(**tokenizer(text, return_tensors="pt"))

# ✅ FAST: Process in batches
batch_size = 32
for i in range(0, len(texts), batch_size):
    batch = texts[i:i+batch_size]
    inputs = tokenizer(batch, padding=True, truncation=True, return_tensors="pt")
    outputs = model(**inputs)

Optimization 2: Mixed Precision (AMP)

from torch.cuda.amp import autocast, GradScaler

scaler = GradScaler()

for batch in dataloader:
    optimizer.zero_grad()

    # Forward pass in mixed precision
    with autocast():
        outputs = model(**batch)
        loss = outputs.loss

    # Backward pass with scaled gradients
    scaler.scale(loss).backward()
    scaler.step(optimizer)
    scaler.update()

Optimization 3: ONNX Export

from transformers import AutoModelForSequenceClassification, AutoTokenizer
from optimum.onnxruntime import ORTModelForSequenceClassification

# Export to ONNX
model = AutoModelForSequenceClassification.from_pretrained("bert-base-uncased")
model.save_pretrained("./onnx-model", export=True)

# Load ONNX model (faster inference)
ort_model = ORTModelForSequenceClassification.from_pretrained("./onnx-model")

# Inference (2-3x faster)
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
inputs = tokenizer("Hello world", return_tensors="pt")
outputs = ort_model(**inputs)

Optimization 4: Dynamic Quantization

import torch

# Quantize model to int8
quantized_model = torch.quantization.quantize_dynamic(
    model,
    {torch.nn.Linear},  # Quantize Linear layers
    dtype=torch.qint8
)

# 4x smaller model, 2-3x faster inference on CPU

Part 6: Common Issues & Solutions

Issue 1: CUDA Out of Memory

Problem: RuntimeError: CUDA out of memory

Solutions:

# Solution 1: Reduce batch size
training_args = TrainingArguments(
    per_device_train_batch_size=8,  # Was 32
    gradient_accumulation_steps=4,  # Effective batch = 8*4 = 32
)

# Solution 2: Use gradient checkpointing
model.gradient_checkpointing_enable()

# Solution 3: Use 8-bit model
from transformers import BitsAndBytesConfig
quantization_config = BitsAndBytesConfig(load_in_8bit=True)
model = AutoModel.from_pretrained("model-name", quantization_config=quantization_config)

# Solution 4: Clear cache
import torch
torch.cuda.empty_cache()

Issue 2: Slow Tokenization

Problem: Tokenization is bottleneck

Solutions:

# Solution 1: Use fast tokenizers
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased", use_fast=True)

# Solution 2: Tokenize dataset once, cache it
tokenized_dataset = dataset.map(
    tokenize_function,
    batched=True,
    num_proc=4,  # Parallel processing
    remove_columns=dataset.column_names,
    load_from_cache_file=True  # Cache results
)

# Solution 3: Use larger batches for tokenization
tokenizer(
    texts,
    padding=True,
    truncation=True,
    max_length=512,
    return_tensors="pt",
    batched=True,  # Process multiple texts at once
    batch_size=1000
)

Issue 3: Inconsistent Results

Problem: Model outputs different results for same input

Solution:

# Set seeds for reproducibility
import random
import numpy as np
import torch

def set_seed(seed=42):
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False

set_seed(42)

# Disable dropout during inference
model.eval()

# Use deterministic generation
outputs = model.generate(
    inputs,
    do_sample=False,  # Greedy decoding
    # OR
    do_sample=True,
    temperature=1.0,
    top_k=50,
    seed=42  # For sampling
)

Issue 4: Attention Mask Errors

Problem: IndexError: index out of range in self

Solution:

# ✅ ALWAYS provide attention mask
tokens = tokenizer(
    text,
    padding=True,
    truncation=True,
    return_tensors="pt",
    return_attention_mask=True  # Explicit (usually default)
)

# Use it in model forward
outputs = model(
    input_ids=tokens['input_ids'],
    attention_mask=tokens['attention_mask']  # Don't forget this!
)

# For custom padding
attention_mask = (input_ids != tokenizer.pad_token_id).long()

Part 7: Model-Specific Patterns

GPT Models (Decoder-Only)

from transformers import GPT2LMHeadModel, GPT2Tokenizer

model = GPT2LMHeadModel.from_pretrained("gpt2")
tokenizer = GPT2Tokenizer.from_pretrained("gpt2")

# Set pad token (GPT doesn't have one by default)
tokenizer.pad_token = tokenizer.eos_token

# Generation
input_text = "The future of AI is"
inputs = tokenizer(input_text, return_tensors="pt")

outputs = model.generate(
    **inputs,
    max_new_tokens=50,
    num_beams=5,  # Beam search
    early_stopping=True,
    no_repeat_ngram_size=2,  # Prevent repetition
    temperature=0.8,
    top_p=0.9
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

T5 Models (Encoder-Decoder)

from transformers import T5ForConditionalGeneration, T5Tokenizer

model = T5ForConditionalGeneration.from_pretrained("t5-small")
tokenizer = T5Tokenizer.from_pretrained("t5-small")

# T5 expects task prefix
input_text = "translate English to German: How are you?"
inputs = tokenizer(input_text, return_tensors="pt")

outputs = model.generate(
    **inputs,
    max_length=50
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))
# "Wie geht es dir?"

BERT Models (Encoder-Only)

from transformers import BertForMaskedLM, BertTokenizer

model = BertForMaskedLM.from_pretrained("bert-base-uncased")
tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")

# Masked language modeling
text = "Paris is the [MASK] of France."
inputs = tokenizer(text, return_tensors="pt")

# Get predictions for [MASK]
outputs = model(**inputs)
mask_token_index = torch.where(inputs["input_ids"] == tokenizer.mask_token_id)[1]
mask_token_logits = outputs.logits[0, mask_token_index, :]

# Top 5 predictions
top_5_tokens = torch.topk(mask_token_logits, 5, dim=1).indices[0].tolist()
for token in top_5_tokens:
    print(tokenizer.decode([token]))
# capital, city, center, heart, ...

Part 8: Production Deployment

FastAPI Serving Pattern

from fastapi import FastAPI
from transformers import pipeline
from pydantic import BaseModel
import uvicorn

app = FastAPI()

# Load model once at startup
classifier = pipeline("text-classification", model="distilbert-base-uncased-finetuned-sst-2-english")

class TextInput(BaseModel):
    text: str

@app.post("/classify")
async def classify_text(input: TextInput):
    result = classifier(input.text)[0]
    return {
        "label": result['label'],
        "confidence": result['score']
    }

if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=8000)

Batch Inference Optimization

import asyncio
from typing import List

class BatchPredictor:
    def __init__(self, model, tokenizer, max_batch_size=32):
        self.model = model
        self.tokenizer = tokenizer
        self.max_batch_size = max_batch_size
        self.queue = []
        self.lock = asyncio.Lock()

    async def predict(self, text: str):
        async with self.lock:
            future = asyncio.Future()
            self.queue.append((text, future))

            if len(self.queue) >= self.max_batch_size:
                await self._process_batch()

        return await future

    async def _process_batch(self):
        if not self.queue:
            return

        texts, futures = zip(*self.queue)
        self.queue = []

        # Process batch
        inputs = self.tokenizer(list(texts), padding=True, truncation=True, return_tensors="pt")
        outputs = self.model(**inputs)
        results = outputs.logits.argmax(dim=-1).tolist()

        # Return results
        for future, result in zip(futures, results):
            future.set_result(result)

Quick Decision Trees

"Which model should I use?"

*Hugging Face Transformers v1.1 - Enhanced*

## 🔄 Workflow

> **Kaynak:** [Hugging Face Course](https://huggingface.co/course) & [Production Guide](https://huggingface.co/docs/transformers/performance)

### Aşama 1: Model Selection
- [ ] **Task**: Görevine en uygun mimariyi seç (Encoder: classification, Decoder: generation).
- [ ] **License**: Modelin ticari kullanım izni (Apache 2.0 vs Llama Community) var mı?
- [ ] **Size**: Parametre sayısı vs performans dengesini kur (7B genellikle yeterli).

### Aşama 2: Optimization pipeline
- [ ] **Quantization**: Inference için 4-bit / 8-bit quantization (BitsAndBytes) kullan.
- [ ] **Batching**: Tek tek değil, batch halinde process et (GPU verimi).
- [ ] **Format**: Production için ONNX veya TensorRT formatına çevir.

### Aşama 3: Deployment
- [ ] **Cache**: Model ağırlıklarını ve tokenizer'ı docker image içine bake etme, volume kullan.
- [ ] **Token Limits**: Context window sınırını aşan inputlar için strateji belirle (chunking).

### Kontrol Noktaları
| Aşama | Doğrulama |
|-------|-----------|
| 1 | Model GPU hafızasına sığıyor mu (OOM hatası)? |
| 2 | Inference süresi (Latency) hedefin altında mı? |
| 3 | Tokenizer ile Model uyumlu mu (aynı vocab)? |
  Classification → BERT, RoBERTa, DeBERTa
  Generation → GPT-2, GPT-Neo, LLaMA
  Translation/Summarization → T5, BART, mT5
  Question Answering → BERT, DeBERTa, RoBERTa

Performance vs Speed?
  Best performance → Large models (355M+ params)
  Balanced → Base models (110M params)
  Fast inference → Distilled models (66M params)

"How should I fine-tune?"

Have full dataset control?
  YES → Full fine-tuning or LoRA
  NO → Few-shot prompting

Dataset size?
  Large (>10K examples) → Full fine-tuning
  Medium (1K-10K) → LoRA or full fine-tuning
  Small (<1K) → LoRA or prompt engineering

Compute available?
  Limited → LoRA (4-bit quantized)
  Moderate → LoRA (8-bit)
  High → Full fine-tuning

Resources

• Hugging Face Docs: https://huggingface.co/docs/transformers/
• Model Hub: https://huggingface.co/models
• PEFT (LoRA): https://huggingface.co/docs/peft/
• Optimum: https://huggingface.co/docs/optimum/
• Datasets: https://huggingface.co/docs/datasets/

Skill version: 1.0.0 Last updated: 2025-10-25 Maintained by: Applied Artificial Intelligence