HomeAboutMeBlogGuest
© 2025 Sejin Cha. All rights reserved.
Built with Next.js, deployed on Vercel
장지원 페이지/
📝
AI Advanced/[LoRA]: Low-Rank Adaptation/
타 코드 리뷰

타 코드 리뷰

 
권태완
Train
성공!
from datasets import load_dataset
from transformers import AutoImageProcessor, AutoModelForImageClassification, TrainingArguments, Trainer
from peft import LoraConfig, get_peft_model
import torch
from torchvision.transforms import (
    Compose,
    Normalize,
    RandomHorizontalFlip,
    RandomResizedCrop,
    ToTensor,
)
import logging

# Initialize logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

# Load CIFAR-10 dataset with 8 classes
def load_cifar10_subset(num_classes=8):
    # Load CIFAR-10 dataset
    dataset = load_dataset("cifar10")

    # Select a subset of classes
    selected_classes = list(range(num_classes))
    dataset = dataset.filter(lambda example: example['label'] in selected_classes)

    labels = dataset["train"].features["label"].names
    label2id = {label: i for i, label in enumerate(labels) if i in selected_classes}
    id2label = {i: label for label, i in label2id.items()}

    return dataset, label2id, id2label

dataset, label2id, id2label = load_cifar10_subset()

# Restrict to 100 samples
dataset['train'] = dataset['train'].select(range(100))

# Image processor and transformations
image_processor = AutoImageProcessor.from_pretrained("google/vit-base-patch16-224-in21k", use_fast=True)

normalize = Normalize(mean=image_processor.image_mean, std=image_processor.image_std)
train_transforms = Compose(
    [
        RandomResizedCrop(image_processor.size["height"]),
        RandomHorizontalFlip(),
        ToTensor(),
        normalize,
    ]
)

def preprocess_train(example_batch):
    example_batch["pixel_values"] = [train_transforms(image) for image in example_batch["img"]]
    return {"pixel_values": example_batch["pixel_values"], "labels": example_batch["label"]}

# Preprocess train dataset
train_dataset = dataset["train"].map(preprocess_train, batched=True, remove_columns=["img", "label"])

# LoRA configuration and model application
model = AutoModelForImageClassification.from_pretrained(
    "google/vit-base-patch16-224-in21k",
    num_labels=len(label2id),
    label2id=label2id,
    id2label=id2label,
)

lora_config = LoraConfig(
    r=16,
    lora_alpha=32,
    target_modules=["attention.query", "attention.key", "attention.value", "intermediate.dense"],
    lora_dropout=0.1,
    bias="none",
    modules_to_save=["classifier"],
)
model = get_peft_model(model, lora_config)

# Training arguments without validation
training_args = TrainingArguments(
    output_dir="./results",
    save_strategy="epoch",
    learning_rate=5e-4,
    per_device_train_batch_size=4,
    num_train_epochs=2,
    weight_decay=0.01,
    logging_steps=10,
    save_total_limit=2,
    load_best_model_at_end=False,
    logging_dir='./logs',
    remove_unused_columns=False,
)

# Data collator definition
def collate_fn(examples):
    pixel_values = torch.stack([torch.tensor(example["pixel_values"]) for example in examples])
    labels = torch.tensor([example["labels"] for example in examples], dtype=torch.long)
    return {"pixel_values": pixel_values, "labels": labels}

# Trainer instance without validation
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    data_collator=collate_fn,
)

# Training
trainer.train()

# 모델과 이미지 프로세서를 로컬에 저장
model.save_pretrained("./trained_model")
image_processor.save_pretrained("./trained_model")
Test
시행착오..
from transformers import AutoImageProcessor, AutoModelForImageClassification
from datasets import load_dataset
from PIL import Image
import torch
import torch.nn.functional as F

# 모델과 이미지 프로세서 로드
model_path = "./trained_model"
model = AutoModelForImageClassification.from_pretrained(model_path, num_labels=8)
image_processor = AutoImageProcessor.from_pretrained(model_path)

# 디바이스 설정
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

# CIFAR-10 데이터셋 로드 (테스트용)
dataset = load_dataset("cifar10")
test_dataset = dataset['test'].select(range(100))  # 테스트 데이터셋에서 100개의 샘플 선택

# 전처리 함수 정의
def preprocess_test(example_batch):
    # 이미지를 텐서로 변환하고, 모델에 필요한 전처리 적용
    example_batch["pixel_values"] = [image_processor(image.convert("RGB"), return_tensors="pt")["pixel_values"] for image in example_batch["img"]]
    return example_batch

# 테스트 데이터셋 전처리
test_dataset = test_dataset.map(preprocess_test, batched=True, remove_columns=["img"])

# 모델 예측 수행
model.eval()
predictions = []
probabilities = []
with torch.no_grad():
    for example in test_dataset:
        pixel_values = torch.cat(example["pixel_values"]).to(device)
        outputs = model(pixel_values)
        logits = outputs.logits
        softmax_probs = F.softmax(logits, dim=-1)
        predicted_class_idx = softmax_probs.argmax(-1).item()
        predicted_prob = softmax_probs.max(-1).values.item() * 100  # 확률을 %로 변환
        predictions.append(predicted_class_idx)
        probabilities.append(predicted_prob)

# 예측된 클래스와 확률 출력
for i, (pred, prob) in enumerate(zip(predictions, probabilities)):
    predicted_class = model.config.id2label[pred]
    print(f"Image {i+1}: Predicted class - {predicted_class}, Probability - {prob:.2f}%")
성공!
from transformers import AutoModelForImageClassification, AutoImageProcessor
from peft import PeftModel, PeftConfig
from datasets import load_dataset
from PIL import Image
import torch
import torch.nn.functional as F

# 모델 경로
model_path = "./trained_model"

# LoRA 어댑터 구성 로드
adapter_config = PeftConfig.from_pretrained(model_path)

# 기본 모델 로드
model = AutoModelForImageClassification.from_pretrained(
    adapter_config.base_model_name_or_path,
    num_labels=8
)

# LoRA 어댑터 적용
model = PeftModel.from_pretrained(model, model_path)

# 이미지 프로세서 로드
image_processor = AutoImageProcessor.from_pretrained(model_path)

# 디바이스 설정
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

# CIFAR-10 데이터셋 로드 (테스트용)
dataset = load_dataset("cifar10")
test_dataset = dataset['test'].select(range(100))  # 테스트 데이터셋에서 100개의 샘플 선택

# 전처리 함수 정의
def preprocess_test(example_batch):
    # 이미지를 텐서로 변환하고, 모델에 필요한 전처리 적용
    example_batch["pixel_values"] = [image_processor(image.convert("RGB"), return_tensors="pt")["pixel_values"] for image in example_batch["img"]]
    return example_batch

# 테스트 데이터셋 전처리
test_dataset = test_dataset.map(preprocess_test, batched=True, remove_columns=["img"])

# 모델 예측 수행
model.eval()
predictions = []
probabilities = []
with torch.no_grad():
    for example in test_dataset:
        # example["pixel_values"]가 리스트인지 확인하고, 텐서로 변환
        if isinstance(example["pixel_values"], list):
            example["pixel_values"] = [torch.tensor(img) for img in example["pixel_values"]]

        # pixel_values가 4D 텐서가 되도록 보장
        pixel_values = torch.stack(example["pixel_values"])

        # 채널 차원이 3이 아니면, 올바르게 변환
        if pixel_values.dim() == 3:
            pixel_values = pixel_values.unsqueeze(1).repeat(1, 3, 1, 1)

        pixel_values = pixel_values.to(device)
        outputs = model(pixel_values)
        logits = outputs.logits
        softmax_probs = F.softmax(logits, dim=-1)
        
        # 배치 내 각 이미지에 대해 예측을 추출
        for i in range(softmax_probs.size(0)):
            predicted_class_idx = softmax_probs[i].argmax(-1).item()
            predicted_prob = softmax_probs[i].max().item() * 100  # 확률을 %로 변환
            predictions.append(predicted_class_idx)
            probabilities.append(predicted_prob)

# 예측된 클래스와 확률 출력
for i, (pred, prob) in enumerate(zip(predictions, probabilities)):
    predicted_class = model.config.id2label[pred]
    print(f"Image {i+1}: Predicted class - {predicted_class}, Probability - {prob:.2f}%")
Some weights of ViTForImageClassification were not initialized from the model checkpoint at google/vit-base-patch16-224-in21k and are newly initialized: ['classifier.bias', 'classifier.weight']
You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
Image 1: Predicted class - LABEL_5, Probability - 16.69%
Image 2: Predicted class - LABEL_5, Probability - 18.50%
Image 3: Predicted class - LABEL_5, Probability - 16.47%
Image 4: Predicted class - LABEL_1, Probability - 16.06%
Image 5: Predicted class - LABEL_1, Probability - 15.07%
Image 6: Predicted class - LABEL_1, Probability - 16.71%
Image 7: Predicted class - LABEL_1, Probability - 15.01%
Image 8: Predicted class - LABEL_1, Probability - 15.40%
Image 9: Predicted class - LABEL_1, Probability - 17.97%
Image 10: Predicted class - LABEL_5, Probability - 15.32%
Image 11: Predicted class - LABEL_1, Probability - 15.12%
Image 12: Predicted class - LABEL_1, Probability - 15.02%
Image 13: Predicted class - LABEL_4, Probability - 15.43%
Image 14: Predicted class - LABEL_5, Probability - 14.43%
Image 15: Predicted class - LABEL_4, Probability - 13.69%
Image 16: Predicted class - LABEL_5, Probability - 15.23%
Image 17: Predicted class - LABEL_5, Probability - 15.03%
Image 18: Predicted class - LABEL_5, Probability - 14.41%
Image 19: Predicted class - LABEL_1, Probability - 19.82%
Image 20: Predicted class - LABEL_1, Probability - 19.55%
Image 21: Predicted class - LABEL_1, Probability - 19.32%
Image 22: Predicted class - LABEL_4, Probability - 15.33%
Image 23: Predicted class - LABEL_4, Probability - 15.29%
Image 24: Predicted class - LABEL_7, Probability - 15.08%
Image 25: Predicted class - LABEL_5, Probability - 16.72%
Image 26: Predicted class - LABEL_5, Probability - 16.90%
Image 27: Predicted class - LABEL_5, Probability - 18.92%
Image 28: Predicted class - LABEL_1, Probability - 16.52%
Image 29: Predicted class - LABEL_1, Probability - 16.34%
Image 30: Predicted class - LABEL_1, Probability - 16.94%
Image 31: Predicted class - LABEL_0, Probability - 14.08%
Image 32: Predicted class - LABEL_3, Probability - 15.22%
Image 33: Predicted class - LABEL_3, Probability - 15.37%
Image 34: Predicted class - LABEL_1, Probability - 17.99%
Image 35: Predicted class - LABEL_1, Probability - 17.53%
Image 36: Predicted class - LABEL_1, Probability - 17.63%
Image 37: Predicted class - LABEL_5, Probability - 17.72%
Image 38: Predicted class - LABEL_5, Probability - 20.68%
Image 39: Predicted class - LABEL_5, Probability - 18.12%
Image 40: Predicted class - LABEL_5, Probability - 14.95%
Image 41: Predicted class - LABEL_5, Probability - 14.93%
Image 42: Predicted class - LABEL_5, Probability - 14.90%
Image 43: Predicted class - LABEL_1, Probability - 16.29%
Image 44: Predicted class - LABEL_1, Probability - 16.76%
Image 45: Predicted class - LABEL_1, Probability - 17.31%
Image 46: Predicted class - LABEL_0, Probability - 14.92%
Image 47: Predicted class - LABEL_2, Probability - 13.54%
Image 48: Predicted class - LABEL_2, Probability - 14.37%
Image 49: Predicted class - LABEL_5, Probability - 20.11%
Image 50: Predicted class - LABEL_5, Probability - 18.14%
Image 51: Predicted class - LABEL_5, Probability - 17.90%
Image 52: Predicted class - LABEL_4, Probability - 14.67%
Image 53: Predicted class - LABEL_2, Probability - 14.06%
Image 54: Predicted class - LABEL_5, Probability - 14.13%
Image 55: Predicted class - LABEL_5, Probability - 14.97%
Image 56: Predicted class - LABEL_5, Probability - 14.73%
Image 57: Predicted class - LABEL_0, Probability - 13.80%
Image 58: Predicted class - LABEL_6, Probability - 16.72%
Image 59: Predicted class - LABEL_6, Probability - 17.31%
Image 60: Predicted class - LABEL_6, Probability - 17.25%
Image 61: Predicted class - LABEL_7, Probability - 15.68%
Image 62: Predicted class - LABEL_7, Probability - 15.94%
Image 63: Predicted class - LABEL_7, Probability - 15.34%
Image 64: Predicted class - LABEL_0, Probability - 14.90%
Image 65: Predicted class - LABEL_0, Probability - 14.73%
Image 66: Predicted class - LABEL_0, Probability - 14.63%
Image 67: Predicted class - LABEL_4, Probability - 13.84%
Image 68: Predicted class - LABEL_3, Probability - 15.44%
Image 69: Predicted class - LABEL_3, Probability - 13.78%
Image 70: Predicted class - LABEL_1, Probability - 17.45%
Image 71: Predicted class - LABEL_1, Probability - 17.38%
Image 72: Predicted class - LABEL_1, Probability - 17.52%
Image 73: Predicted class - LABEL_5, Probability - 15.92%
Image 74: Predicted class - LABEL_5, Probability - 15.08%
Image 75: Predicted class - LABEL_5, Probability - 14.83%
Image 76: Predicted class - LABEL_5, Probability - 16.18%
Image 77: Predicted class - LABEL_2, Probability - 15.25%
Image 78: Predicted class - LABEL_2, Probability - 15.61%
Image 79: Predicted class - LABEL_5, Probability - 18.82%
Image 80: Predicted class - LABEL_5, Probability - 17.00%
Image 81: Predicted class - LABEL_5, Probability - 14.59%
Image 82: Predicted class - LABEL_0, Probability - 15.57%
Image 83: Predicted class - LABEL_0, Probability - 14.68%
Image 84: Predicted class - LABEL_0, Probability - 15.15%
Image 85: Predicted class - LABEL_1, Probability - 17.71%
Image 86: Predicted class - LABEL_1, Probability - 16.98%
Image 87: Predicted class - LABEL_1, Probability - 16.62%
Image 88: Predicted class - LABEL_6, Probability - 15.67%
Image 89: Predicted class - LABEL_5, Probability - 13.45%
Image 90: Predicted class - LABEL_5, Probability - 14.23%
Image 91: Predicted class - LABEL_4, Probability - 14.12%
Image 92: Predicted class - LABEL_5, Probability - 15.24%
Image 93: Predicted class - LABEL_4, Probability - 14.80%
Image 94: Predicted class - LABEL_5, Probability - 16.71%
Image 95: Predicted class - LABEL_5, Probability - 15.63%
Image 96: Predicted class - LABEL_5, Probability - 15.08%
Image 97: Predicted class - LABEL_5, Probability - 14.44%
Image 98: Predicted class - LABEL_0, Probability - 14.00%
Image 99: Predicted class - LABEL_5, Probability - 14.10%
Image 100: Predicted class - LABEL_5, Probability - 14.83%
Image 101: Predicted class - LABEL_5, Probability - 14.32%
Image 102: Predicted class - LABEL_5, Probability - 14.56%
Image 103: Predicted class - LABEL_1, Probability - 18.00%
Image 104: Predicted class - LABEL_1, Probability - 18.18%
Image 105: Predicted class - LABEL_1, Probability - 18.72%
Image 106: Predicted class - LABEL_2, Probability - 15.25%
Image 107: Predicted class - LABEL_2, Probability - 15.47%
Image 108: Predicted class - LABEL_2, Probability - 15.34%
Image 109: Predicted class - LABEL_5, Probability - 16.51%
Image 110: Predicted class - LABEL_5, Probability - 16.51%
Image 111: Predicted class - LABEL_4, Probability - 15.79%
Image 112: Predicted class - LABEL_1, Probability - 18.80%
Image 113: Predicted class - LABEL_1, Probability - 18.80%
Image 114: Predicted class - LABEL_1, Probability - 18.80%
Image 115: Predicted class - LABEL_1, Probability - 18.32%
Image 116: Predicted class - LABEL_1, Probability - 15.84%
Image 117: Predicted class - LABEL_1, Probability - 16.30%
Image 118: Predicted class - LABEL_5, Probability - 16.54%
Image 119: Predicted class - LABEL_5, Probability - 15.66%
Image 120: Predicted class - LABEL_5, Probability - 14.68%
Image 121: Predicted class - LABEL_5, Probability - 15.68%
Image 122: Predicted class - LABEL_5, Probability - 15.21%
Image 123: Predicted class - LABEL_5, Probability - 14.93%
Image 124: Predicted class - LABEL_5, Probability - 15.62%
Image 125: Predicted class - LABEL_5, Probability - 14.93%
Image 126: Predicted class - LABEL_2, Probability - 14.60%
Image 127: Predicted class - LABEL_5, Probability - 17.00%
Image 128: Predicted class - LABEL_5, Probability - 18.82%
Image 129: Predicted class - LABEL_5, Probability - 16.16%
Image 130: Predicted class - LABEL_5, Probability - 15.23%
Image 131: Predicted class - LABEL_5, Probability - 14.41%
Image 132: Predicted class - LABEL_5, Probability - 14.69%
Image 133: Predicted class - LABEL_0, Probability - 15.73%
Image 134: Predicted class - LABEL_0, Probability - 15.21%
Image 135: Predicted class - LABEL_0, Probability - 15.58%
Image 136: Predicted class - LABEL_1, Probability - 17.96%
Image 137: Predicted class - LABEL_1, Probability - 18.33%
Image 138: Predicted class - LABEL_1, Probability - 18.60%
Image 139: Predicted class - LABEL_5, Probability - 17.68%
Image 140: Predicted class - LABEL_5, Probability - 16.22%
Image 141: Predicted class - LABEL_5, Probability - 15.40%
Image 142: Predicted class - LABEL_3, Probability - 16.41%
Image 143: Predicted class - LABEL_3, Probability - 17.52%
Image 144: Predicted class - LABEL_3, Probability - 14.86%
Image 145: Predicted class - LABEL_5, Probability - 14.37%
Image 146: Predicted class - LABEL_7, Probability - 14.13%
Image 147: Predicted class - LABEL_1, Probability - 13.84%
Image 148: Predicted class - LABEL_5, Probability - 13.84%
Image 149: Predicted class - LABEL_4, Probability - 14.05%
Image 150: Predicted class - LABEL_5, Probability - 13.83%
Image 151: Predicted class - LABEL_1, Probability - 15.27%
Image 152: Predicted class - LABEL_1, Probability - 15.08%
Image 153: Predicted class - LABEL_1, Probability - 14.96%
Image 154: Predicted class - LABEL_5, Probability - 13.37%
Image 155: Predicted class - LABEL_0, Probability - 14.19%
Image 156: Predicted class - LABEL_0, Probability - 14.34%
Image 157: Predicted class - LABEL_0, Probability - 16.42%
Image 158: Predicted class - LABEL_0, Probability - 16.11%
Image 159: Predicted class - LABEL_0, Probability - 16.97%
Image 160: Predicted class - LABEL_5, Probability - 15.74%
Image 161: Predicted class - LABEL_5, Probability - 15.72%
Image 162: Predicted class - LABEL_5, Probability - 15.26%
Image 163: Predicted class - LABEL_1, Probability - 14.48%
Image 164: Predicted class - LABEL_1, Probability - 14.36%
Image 165: Predicted class - LABEL_1, Probability - 14.05%
Image 166: Predicted class - LABEL_5, Probability - 17.24%
Image 167: Predicted class - LABEL_5, Probability - 17.62%
Image 168: Predicted class - LABEL_5, Probability - 17.05%
Image 169: Predicted class - LABEL_5, Probability - 15.83%
Image 170: Predicted class - LABEL_5, Probability - 15.58%
Image 171: Predicted class - LABEL_5, Probability - 15.77%
Image 172: Predicted class - LABEL_5, Probability - 15.64%
Image 173: Predicted class - LABEL_5, Probability - 17.23%
Image 174: Predicted class - LABEL_5, Probability - 18.52%
Image 175: Predicted class - LABEL_5, Probability - 18.04%
Image 176: Predicted class - LABEL_5, Probability - 16.58%
Image 177: Predicted class - LABEL_5, Probability - 16.42%
Image 178: Predicted class - LABEL_5, Probability - 17.10%
Image 179: Predicted class - LABEL_5, Probability - 17.33%
Image 180: Predicted class - LABEL_5, Probability - 16.76%
Image 181: Predicted class - LABEL_7, Probability - 14.55%
Image 182: Predicted class - LABEL_5, Probability - 14.79%
Image 183: Predicted class - LABEL_5, Probability - 14.23%
Image 184: Predicted class - LABEL_5, Probability - 15.75%
Image 185: Predicted class - LABEL_5, Probability - 17.21%
Image 186: Predicted class - LABEL_2, Probability - 15.09%
Image 187: Predicted class - LABEL_5, Probability - 14.06%
Image 188: Predicted class - LABEL_5, Probability - 14.27%
Image 189: Predicted class - LABEL_5, Probability - 15.67%
Image 190: Predicted class - LABEL_1, Probability - 15.00%
Image 191: Predicted class - LABEL_1, Probability - 16.09%
Image 192: Predicted class - LABEL_1, Probability - 16.22%
Image 193: Predicted class - LABEL_5, Probability - 14.73%
Image 194: Predicted class - LABEL_5, Probability - 15.23%
Image 195: Predicted class - LABEL_5, Probability - 16.24%
Image 196: Predicted class - LABEL_2, Probability - 14.11%
Image 197: Predicted class - LABEL_2, Probability - 15.59%
Image 198: Predicted class - LABEL_2, Probability - 14.51%
Image 199: Predicted class - LABEL_1, Probability - 15.93%
Image 200: Predicted class - LABEL_1, Probability - 16.64%
Image 201: Predicted class - LABEL_1, Probability - 17.48%
Image 202: Predicted class - LABEL_5, Probability - 13.59%
Image 203: Predicted class - LABEL_5, Probability - 14.03%
Image 204: Predicted class - LABEL_1, Probability - 13.58%
Image 205: Predicted class - LABEL_5, Probability - 14.78%
Image 206: Predicted class - LABEL_4, Probability - 14.17%
Image 207: Predicted class - LABEL_1, Probability - 15.50%
Image 208: Predicted class - LABEL_5, Probability - 14.10%
Image 209: Predicted class - LABEL_5, Probability - 15.47%
Image 210: Predicted class - LABEL_5, Probability - 15.95%
Image 211: Predicted class - LABEL_3, Probability - 15.22%
Image 212: Predicted class - LABEL_0, Probability - 14.31%
Image 213: Predicted class - LABEL_0, Probability - 15.26%
Image 214: Predicted class - LABEL_1, Probability - 14.60%
Image 215: Predicted class - LABEL_5, Probability - 14.41%
Image 216: Predicted class - LABEL_5, Probability - 14.67%
Image 217: Predicted class - LABEL_0, Probability - 14.52%
Image 218: Predicted class - LABEL_0, Probability - 14.13%
Image 219: Predicted class - LABEL_0, Probability - 14.06%
Image 220: Predicted class - LABEL_1, Probability - 13.44%
Image 221: Predicted class - LABEL_4, Probability - 13.75%
Image 222: Predicted class - LABEL_4, Probability - 14.34%
Image 223: Predicted class - LABEL_0, Probability - 15.77%
Image 224: Predicted class - LABEL_0, Probability - 15.46%
Image 225: Predicted class - LABEL_0, Probability - 16.73%
Image 226: Predicted class - LABEL_2, Probability - 13.94%
Image 227: Predicted class - LABEL_2, Probability - 16.36%
Image 228: Predicted class - LABEL_2, Probability - 12.98%
Image 229: Predicted class - LABEL_1, Probability - 15.90%
Image 230: Predicted class - LABEL_1, Probability - 17.01%
Image 231: Predicted class - LABEL_1, Probability - 16.94%
Image 232: Predicted class - LABEL_5, Probability - 15.34%
Image 233: Predicted class - LABEL_5, Probability - 16.89%
Image 234: Predicted class - LABEL_5, Probability - 14.48%
Image 235: Predicted class - LABEL_5, Probability - 16.90%
Image 236: Predicted class - LABEL_5, Probability - 16.70%
Image 237: Predicted class - LABEL_5, Probability - 15.56%
Image 238: Predicted class - LABEL_5, Probability - 14.56%
Image 239: Predicted class - LABEL_5, Probability - 14.56%
Image 240: Predicted class - LABEL_5, Probability - 14.56%
Image 241: Predicted class - LABEL_1, Probability - 14.09%
Image 242: Predicted class - LABEL_1, Probability - 14.00%
Image 243: Predicted class - LABEL_1, Probability - 14.17%
Image 244: Predicted class - LABEL_5, Probability - 14.45%
Image 245: Predicted class - LABEL_0, Probability - 14.58%
Image 246: Predicted class - LABEL_0, Probability - 14.48%
Image 247: Predicted class - LABEL_1, Probability - 18.88%
Image 248: Predicted class - LABEL_1, Probability - 18.91%
Image 249: Predicted class - LABEL_1, Probability - 20.98%
Image 250: Predicted class - LABEL_7, Probability - 14.33%
Image 251: Predicted class - LABEL_7, Probability - 14.33%
Image 252: Predicted class - LABEL_7, Probability - 14.33%
Image 253: Predicted class - LABEL_2, Probability - 16.27%
Image 254: Predicted class - LABEL_2, Probability - 16.36%
Image 255: Predicted class - LABEL_2, Probability - 16.45%
Image 256: Predicted class - LABEL_5, Probability - 17.19%
Image 257: Predicted class - LABEL_5, Probability - 16.91%
Image 258: Predicted class - LABEL_5, Probability - 15.71%
Image 259: Predicted class - LABEL_5, Probability - 15.50%
Image 260: Predicted class - LABEL_5, Probability - 15.60%
Image 261: Predicted class - LABEL_5, Probability - 14.15%
Image 262: Predicted class - LABEL_4, Probability - 14.31%
Image 263: Predicted class - LABEL_4, Probability - 14.19%
Image 264: Predicted class - LABEL_7, Probability - 14.17%
Image 265: Predicted class - LABEL_2, Probability - 14.59%
Image 266: Predicted class - LABEL_2, Probability - 14.58%
Image 267: Predicted class - LABEL_2, Probability - 13.59%
Image 268: Predicted class - LABEL_1, Probability - 17.45%
Image 269: Predicted class - LABEL_1, Probability - 18.66%
Image 270: Predicted class - LABEL_1, Probability - 17.22%
Image 271: Predicted class - LABEL_7, Probability - 15.80%
Image 272: Predicted class - LABEL_0, Probability - 16.00%
Image 273: Predicted class - LABEL_0, Probability - 16.48%
Image 274: Predicted class - LABEL_5, Probability - 15.45%
Image 275: Predicted class - LABEL_5, Probability - 15.73%
Image 276: Predicted class - LABEL_5, Probability - 15.51%
Image 277: Predicted class - LABEL_1, Probability - 14.73%
Image 278: Predicted class - LABEL_3, Probability - 14.72%
Image 279: Predicted class - LABEL_0, Probability - 13.74%
Image 280: Predicted class - LABEL_7, Probability - 15.14%
Image 281: Predicted class - LABEL_7, Probability - 14.88%
Image 282: Predicted class - LABEL_3, Probability - 14.69%
Image 283: Predicted class - LABEL_5, Probability - 17.83%
Image 284: Predicted class - LABEL_5, Probability - 14.83%
Image 285: Predicted class - LABEL_5, Probability - 15.09%
Image 286: Predicted class - LABEL_2, Probability - 13.43%
Image 287: Predicted class - LABEL_1, Probability - 13.75%
Image 288: Predicted class - LABEL_1, Probability - 13.51%
Image 289: Predicted class - LABEL_5, Probability - 13.79%
Image 290: Predicted class - LABEL_5, Probability - 13.85%
Image 291: Predicted class - LABEL_6, Probability - 15.14%
Image 292: Predicted class - LABEL_7, Probability - 14.43%
Image 293: Predicted class - LABEL_7, Probability - 14.21%
Image 294: Predicted class - LABEL_5, Probability - 14.19%
Image 295: Predicted class - LABEL_1, Probability - 15.93%
Image 296: Predicted class - LABEL_1, Probability - 16.06%
Image 297: Predicted class - LABEL_1, Probability - 15.85%
Image 298: Predicted class - LABEL_5, Probability - 14.13%
Image 299: Predicted class - LABEL_7, Probability - 14.11%
Image 300: Predicted class - LABEL_4, Probability - 14.38%
성공! (2)
from transformers import AutoModelForImageClassification, AutoImageProcessor
from peft import PeftModel, PeftConfig
from datasets import load_dataset
import torch
import torch.nn.functional as F

# 모델 경로
model_path = "./trained_model"

# LoRA 어댑터 구성 로드
adapter_config = PeftConfig.from_pretrained(model_path)

# 기본 모델 로드
model = AutoModelForImageClassification.from_pretrained(
    adapter_config.base_model_name_or_path,
    num_labels=8  # CIFAR-10은 10개의 클래스를 가짐
)

# LoRA 어댑터 적용
model = PeftModel.from_pretrained(model, model_path)

# 이미지 프로세서 로드
image_processor = AutoImageProcessor.from_pretrained(model_path)

# 디바이스 설정
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

# CIFAR-10 데이터셋 로드 (테스트용)
dataset = load_dataset("cifar10")
test_dataset = dataset['test'].select(range(100))  # 테스트 데이터셋에서 100개의 샘플 선택

# 전처리 함수 정의
def preprocess_test(example_batch):
    # 이미지를 텐서로 변환하고, 모델에 필요한 전처리 적용
    example_batch["pixel_values"] = [image_processor(image.convert("RGB"), return_tensors="pt")["pixel_values"] for image in example_batch["img"]]
    return example_batch

# 테스트 데이터셋 전처리
test_dataset = test_dataset.map(preprocess_test, batched=True, remove_columns=["img"])

# 모델 예측 수행 및 정확도 계산
model.eval()
correct_predictions = 0
total_predictions = 0

with torch.no_grad():
    for example in test_dataset:
        # example["pixel_values"]가 리스트인지 확인하고, 텐서로 변환
        if isinstance(example["pixel_values"], list):
            example["pixel_values"] = [torch.tensor(img) for img in example["pixel_values"]]

        # pixel_values가 4D 텐서가 되도록 보장
        pixel_values = torch.stack(example["pixel_values"])

        # 채널 차원이 3이 아니면, 올바르게 변환
        if pixel_values.dim() == 3:
            pixel_values = pixel_values.unsqueeze(1).repeat(1, 3, 1, 1)

        pixel_values = pixel_values.to(device)
        outputs = model(pixel_values)
        logits = outputs.logits
        predictions = torch.argmax(logits, dim=-1)

        # 실제 라벨과 예측 라벨 비교
        labels = torch.tensor(example['label']).to(device)

        # labels가 스칼라 값일 경우 1차원 텐서로 변환
        if labels.dim() == 0:
            labels = labels.unsqueeze(0)

        correct_predictions += (predictions == labels).sum().item()
        total_predictions += labels.size(0)

# 정확도 계산 및 출력
accuracy = (correct_predictions / total_predictions) * 100
print(f"Correct Predictions: {correct_predictions}/{total_predictions}")
print(f"Accuracy: {accuracy:.2f}%")