2024/7/21 - ViT

ViT를 PEFT형식으로 전이학습하기

사용한 Dataset: huggingfaceFastJobs/Visual_Emotional_Analysis · Datasets at Hugging Face

FastJobs/Visual_Emotional_Analysis · Datasets at Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface

사용한 모델: huggingfacegoogle/vit-base-patch16-224-in21k · Hugging Face

google/vit-base-patch16-224-in21k · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface

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 (
    CenterCrop,
    Compose,
    Normalize,
    RandomHorizontalFlip,
    RandomResizedCrop,
    Resize,
    ToTensor,
)
from PIL import Image
import logging

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

# Load dataset
dataset = load_dataset("FastJobs/Visual_Emotional_Analysis")

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

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

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.convert("RGB")) for image in example_batch["image"]]
    return {"pixel_values": example_batch["pixel_values"], "labels": example_batch["label"]}

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

# LoRA configuration and model application
model = AutoModelForImageClassification.from_pretrained(
    "google/vit-base-patch16-224-in21k",
    num_labels=len(labels),  # 올바른 클래스 수 설정
    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',
)

# 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")

Fast image processor class <class 'transformers.models.vit.image_processing_vit_fast.ViTImageProcessorFast'> is available for this model. Using slow image processor class. To use the fast image processor class set `use_fast=True`.
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.

테스트

from transformers import AutoImageProcessor, AutoModelForImageClassification
from PIL import Image
import torch
import os

# 로그 테스트 - logging이라는 라이브러리가 작동하지 않아 print문으로 대체하였음!
print("로그 테스트 - 이 메시지가 출력되면 로그 설정이 정상적으로 작동하는 것입니다.")

# 모델과 이미지 프로세서 로드
try:
    print("모델 로드 중...")
    model_path = "./trained_model"
    if not os.path.exists(model_path):
        print("모델 경로가 존재하지 않습니다:", model_path)
        raise FileNotFoundError(f"모델 경로가 존재하지 않습니다: {model_path}")

    model = AutoModelForImageClassification.from_pretrained(model_path, num_labels=8)
    #num_labels를 설정하지 않으면 오류가 생긴다!!!
    image_processor = AutoImageProcessor.from_pretrained(model_path)
    print("모델 및 이미지 프로세서 로드 완료")
except Exception as e:
    print("모델 로드 실패:", e)
    raise e

# 이미지 파일 경로
uploaded_image_path = "./content/amusement.jpg"

# 이미지 로드 및 전처리
try:
    print("이미지 로드 중...")
    if not os.path.exists(uploaded_image_path):
        print("이미지 파일이 존재하지 않습니다:", uploaded_image_path)
        raise FileNotFoundError(f"이미지 파일이 존재하지 않습니다: {uploaded_image_path}")

    image = Image.open(uploaded_image_path)
    encoding = image_processor(images=image.convert("RGB"), return_tensors="pt")
    print("이미지 전처리 완료")
except Exception as e:
    print("이미지 로드 및 전처리 실패:", e)
    raise e

# 모델을 동일한 디바이스에 이동
try:
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    print(f"디바이스: {device}")
    model.to(device)
    encoding = {k: v.to(device) for k, v in encoding.items()}
    print("디바이스로 이동된 encoding:", encoding)
except Exception as e:
    print("디바이스 이동 실패:", e)
    raise e

# 모델 예측 수행
try:
    print("모델 예측 수행 중...")
    model.eval()
    with torch.no_grad():
        outputs = model(**encoding)
        logits = outputs.logits
    print("예측 완료")
except Exception as e:
    print("모델 예측 실패:", e)
    raise e

# 예측된 클래스 출력
try:
    predicted_class_idx = logits.argmax(-1).item()
    predicted_class = model.config.id2label[predicted_class_idx]
    print("Predicted class:", predicted_class)
except Exception as e:
    print("예측된 클래스 출력 실패:", e)
    raise e

로그 테스트 - 이 메시지가 출력되면 로그 설정이 정상적으로 작동하는 것입니다.
모델 로드 중...

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.

모델 및 이미지 프로세서 로드 완료
이미지 로드 중...
이미지 전처리 완료
디바이스: cuda
디바이스로 이동된 encoding: {'pixel_values': tensor([[[[0.7882, 0.8275, 0.8510,  ..., 0.8667, 0.8745, 0.9059],
          [0.8118, 0.8510, 0.8431,  ..., 0.8510, 0.8824, 0.9059],
          [0.8196, 0.8588, 0.8745,  ..., 0.8431, 0.8667, 0.8902],
          ...,
          [0.8118, 0.8118, 0.7804,  ..., 0.6784, 0.6627, 0.6314],
          [0.8510, 0.8275, 0.8353,  ..., 0.7176, 0.7176, 0.7255],
          [0.8431, 0.8275, 0.8353,  ..., 0.7333, 0.7176, 0.7176]],

         [[0.7725, 0.7961, 0.7961,  ..., 0.8588, 0.8510, 0.8510],
          [0.7961, 0.8196, 0.7961,  ..., 0.8431, 0.8510, 0.8588],
          [0.8039, 0.8353, 0.8196,  ..., 0.8431, 0.8353, 0.8353],
          ...,
          [0.8275, 0.8275, 0.7961,  ..., 0.7333, 0.7255, 0.6941],
          [0.8667, 0.8353, 0.8510,  ..., 0.7725, 0.7647, 0.7804],
          [0.8510, 0.8353, 0.8431,  ..., 0.7804, 0.7647, 0.7647]],

         [[0.6549, 0.6863, 0.6941,  ..., 0.7725, 0.7647, 0.7725],
          [0.6784, 0.7020, 0.6863,  ..., 0.7490, 0.7647, 0.7725],
          [0.6863, 0.7176, 0.7176,  ..., 0.7490, 0.7490, 0.7569],
          ...,
          [0.6549, 0.6549, 0.6235,  ..., 0.4588, 0.4431, 0.4196],
          [0.7098, 0.6784, 0.6941,  ..., 0.5529, 0.5529, 0.5608],
          [0.7098, 0.6941, 0.7020,  ..., 0.5608, 0.5529, 0.5451]]]],
       device='cuda:0')}
모델 예측 수행 중...
예측 완료
Predicted class: LABEL_7

로그 테스트 - 이 메시지가 출력되면 로그 설정이 정상적으로 작동하는 것입니다.
모델 로드 중...

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.

모델 및 이미지 프로세서 로드 완료
이미지 로드 중...
이미지 전처리 완료
디바이스: cuda
디바이스로 이동된 encoding: {'pixel_values': tensor([[[[ 0.1373,  0.0980,  0.0510,  ...,  0.3020,  0.2863,  0.2627],
          [ 0.0510, -0.0039, -0.0431,  ..., -0.0353, -0.0510, -0.0745],
          [ 0.0039, -0.0431, -0.0745,  ..., -0.1529, -0.1529, -0.1686],
          ...,
          [ 0.4275,  0.5059,  0.5216,  ...,  0.4196,  0.4667,  0.5137],
          [ 0.4824,  0.5059,  0.4902,  ...,  0.4353,  0.4824,  0.5294],
          [ 0.4980,  0.4745,  0.4588,  ...,  0.4353,  0.4824,  0.5294]],

         [[ 0.2314,  0.2000,  0.1608,  ...,  0.3098,  0.2941,  0.2706],
          [ 0.0980,  0.0588,  0.0353,  ...,  0.0353,  0.0196, -0.0039],
          [ 0.0431,  0.0196,  0.0039,  ..., -0.1059, -0.1059, -0.1216],
          ...,
          [ 0.5294,  0.6078,  0.6392,  ...,  0.4824,  0.5529,  0.6078],
          [ 0.5922,  0.6314,  0.6235,  ...,  0.5216,  0.5843,  0.6392],
          [ 0.6157,  0.6078,  0.6078,  ...,  0.5373,  0.6000,  0.6627]],

         [[ 0.1529,  0.1294,  0.1059,  ...,  0.2784,  0.2627,  0.2392],
          [-0.0039, -0.0510, -0.0745,  ...,  0.0118, -0.0039, -0.0275],
          [-0.0510, -0.0902, -0.1294,  ..., -0.2235, -0.2235, -0.2392],
          ...,
          [ 0.6863,  0.7804,  0.8431,  ...,  0.7020,  0.7882,  0.8431],
          [ 0.7882,  0.8353,  0.8353,  ...,  0.7333,  0.8118,  0.8667],
          [ 0.8510,  0.8353,  0.8275,  ...,  0.7490,  0.8275,  0.8824]]]],
       device='cuda:0')}
모델 예측 수행 중...
예측 완료
Predicted class: LABEL_1

최종 파일 업로드!!!

Google Colab

Google Colab