PyTorch Lightning

pip install lightning -q
pip install wandb -qU

import lightning.pytorch as pl

# 즐겨 사용하는 기계학습 트래킹 툴
from lightning.pytorch.loggers import WandbLogger

import torch
from torch import nn
from torch.nn import functional as F
from torch.utils.data import random_split, DataLoader

from torchmetrics import Accuracy

from torchvision import transforms
from torchvision.datasets import CIFAR10

import wandb

wandb.login()

class CIFAR10DataModule(pl.LightningDataModule):
    def __init__(self, batch_size, data_dir: str = './'):
        super().__init__()
        self.data_dir = data_dir
        self.batch_size = batch_size

        self.transform = transforms.Compose([
            transforms.ToTensor(),
            transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
        ])
        
        self.num_classes = 10
    
    def prepare_data(self):
        CIFAR10(self.data_dir, train=True, download=True)
        CIFAR10(self.data_dir, train=False, download=True)
    
    def setup(self, stage=None):
        # 데이터로더에서 사용할 트레이닝/검증 데이터셋 할당
        if stage == 'fit' or stage is None:
            cifar_full = CIFAR10(self.data_dir, train=True, transform=self.transform)
            self.cifar_train, self.cifar_val = random_split(cifar_full, [45000, 5000])

        # 데이터로더에서 사용할 테스트 데이터셋 할당
        if stage == 'test' or stage is None:
            self.cifar_test = CIFAR10(self.data_dir, train=False, transform=self.transform)
    
    def train_dataloader(self):
        return DataLoader(self.cifar_train, batch_size=self.batch_size, shuffle=True)

    def val_dataloader(self):
        return DataLoader(self.cifar_val, batch_size=self.batch_size)

    def test_dataloader(self):
        return DataLoader(self.cifar_test, batch_size=self.batch_size)

class ImagePredictionLogger(pl.callbacks.Callback):
    def __init__(self, val_samples, num_samples=32):
        super().__init__()
        self.num_samples = num_samples
        self.val_imgs, self.val_labels = val_samples
    
    def on_validation_epoch_end(self, trainer, pl_module):
        # 텐서를 CPU로 이동
        val_imgs = self.val_imgs.to(device=pl_module.device)
        val_labels = self.val_labels.to(device=pl_module.device)
        # 모델 예측값 가져오기
        logits = pl_module(val_imgs)
        preds = torch.argmax(logits, -1)
        # 이미지를 wandb Image로 로그 기록
        trainer.logger.experiment.log({
            "examples":[wandb.Image(x, caption=f"Pred:{pred}, Label:{y}") 
                           for x, pred, y in zip(val_imgs[:self.num_samples], 
                                                 preds[:self.num_samples], 
                                                 val_labels[:self.num_samples])]
            })
        

class LitModel(pl.LightningModule):
    def __init__(self, input_shape, num_classes, learning_rate=2e-4):
        super().__init__()
        
        # 하이퍼파라미터 로그 기록
        self.save_hyperparameters()
        self.learning_rate = learning_rate
        
        self.conv1 = nn.Conv2d(3, 32, 3, 1)
        self.conv2 = nn.Conv2d(32, 32, 3, 1)
        self.conv3 = nn.Conv2d(32, 64, 3, 1)
        self.conv4 = nn.Conv2d(64, 64, 3, 1)

        self.pool1 = torch.nn.MaxPool2d(2)
        self.pool2 = torch.nn.MaxPool2d(2)
        
        n_sizes = self._get_conv_output(input_shape)

        self.fc1 = nn.Linear(n_sizes, 512)
        self.fc2 = nn.Linear(512, 128)
        self.fc3 = nn.Linear(128, num_classes)

        self.accuracy = Accuracy(task='multiclass', num_classes=num_classes)

    # conv 블록에서 Linear 레이어로 들어가는 출력 텐서의 크기를 반환합니다.
    def _get_conv_output(self, shape):
        batch_size = 1
        input = torch.autograd.Variable(torch.rand(batch_size, *shape))

        output_feat = self._forward_features(input) 
        n_size = output_feat.data.view(batch_size, -1).size(1)
        return n_size
        
    # conv 블록에서 특징 텐서를 반환합니다.
    def _forward_features(self, x):
        x = F.relu(self.conv1(x))
        x = self.pool1(F.relu(self.conv2(x)))
        x = F.relu(self.conv3(x))
        x = self.pool2(F.relu(self.conv4(x)))
        return x
    
    # 추론 시에 사용됩니다.
    def forward(self, x):
       x = self._forward_features(x)
       x = x.view(x.size(0), -1)
       x = F.relu(self.fc1(x))
       x = F.relu(self.fc2(x))
       x = F.log_softmax(self.fc3(x), dim=1)
       
       return x
    
    def training_step(self, batch, batch_idx):
        x, y = batch
        logits = self(x)
        loss = F.nll_loss(logits, y)
        
        # 트레이닝 메트릭
        preds = torch.argmax(logits, dim=1)
        acc = self.accuracy(preds, y)
        self.log('train_loss', loss, on_step=True, on_epoch=True, logger=True)
        self.log('train_acc', acc, on_step=True, on_epoch=True, logger=True)
        
        return loss
    
    def validation_step(self, batch, batch_idx):
        x, y = batch
        logits = self(x)
        loss = F.nll_loss(logits, y)

        # 검증 메트릭
        preds = torch.argmax(logits, dim=1)
        acc = self.accuracy(preds, y)
        self.log('val_loss', loss, prog_bar=True)
        self.log('val_acc', acc, prog_bar=True)
        return loss
    
    def test_step(self, batch, batch_idx):
        x, y = batch
        logits = self(x)
        loss = F.nll_loss(logits, y)
        
        # 검증 메트릭
        preds = torch.argmax(logits, dim=1)
        acc = self.accuracy(preds, y)
        self.log('test_loss', loss, prog_bar=True)
        self.log('test_acc', acc, prog_bar=True)
        return loss
    
    def configure_optimizers(self):
        optimizer = torch.optim.Adam(self.parameters(), lr=self.learning_rate)
        return optimizer

dm = CIFAR10DataModule(batch_size=32)
# 데이터로더에 엑세스하려면 prepare_data와 setup을 호출해야 합니다.
dm.prepare_data()
dm.setup()

# 커스텀 ImagePredictionLogger 콜백에서 이미지 예측을 기록하는 데 필요한 샘플입니다.
val_samples = next(iter(dm.val_dataloader()))
val_imgs, val_labels = val_samples[0], val_samples[1]
val_imgs.shape, val_labels.shape

model = LitModel((3, 32, 32), dm.num_classes)

# wandb logger 초기화
wandb_logger = WandbLogger(project='wandb-lightning', job_type='train')

# 콜백 초기화
early_stop_callback = pl.callbacks.EarlyStopping(monitor="val_loss")
checkpoint_callback = pl.callbacks.ModelCheckpoint()

# trainer 초기화
trainer = pl.Trainer(max_epochs=2,
                     logger=wandb_logger,
                     callbacks=[early_stop_callback,
                                ImagePredictionLogger(val_samples),
                                checkpoint_callback],
                     )

# 모델 트레이닝 
trainer.fit(model, dm)

# 별도의 테스트 세트에서 모델 평가 ⚡⚡
trainer.test(dataloaders=dm.test_dataloader())

# wandb run 종료
run.finish()