ResNet 만들기

import torch import torch.nn as nn class BasicBlock(nn.Module): def __init__(self, in_channels, out_channels, kernel_size=3): super(BasicBlock, self).__init__() # 1) 합성곱층 정의 self.c1 = nn.Conv2d( in_channels, out_channels, kernel_size=kernel_size, padding=1 ) self.c2 = nn.Conv2d( out_channels, out_channels, kernel_size=kernel_size, padding=1 ) # 입력과 출력의 채널 수를 맞추기 위한 1×1 합성곱(다운샘플) self.downsample = nn.Conv2d(in_channels, out_channels, kernel_size=1) # 2) 배치 정규화층 정의 self.bn1 = nn.BatchNorm2d(num_features=out_channels) self.bn2 = nn.BatchNorm2d(num_features=out_channels) self.relu = nn.ReLU()

def forward(self, x): # 스킵 커넥션을 위해 초기 입력 저장 x_ = x # ResNet 기본 블록에서 F(x) 부분 x = self.c1(x) x = self.bn1(x) x = self.relu(x) x = self.c2(x) x = self.bn2(x) # 합성곱 결과와 입력의 채널 수를 맞춤 (다운샘플) x_ = self.downsample(x_) # 합성곱층 결과와 저장해 둔 입력을 더해줌 (스킵 커넥션) x += x_ x = self.relu(x) return x

class ResNet(nn.Module): def __init__(self, num_classes=10): super(ResNet, self).__init__() # 기본 블록 self.b1 = BasicBlock(in_channels=3, out_channels=64) self.b2 = BasicBlock(in_channels=64, out_channels=128) self.b3 = BasicBlock(in_channels=128, out_channels=256) # 풀링을 최댓값이 아닌 평균값으로 self.pool = nn.AvgPool2d(kernel_size=2, stride=2) # 분류기 self.fc1 = nn.Linear(in_features=4096, out_features=2048) self.fc2 = nn.Linear(in_features=2048, out_features=512) self.fc3 = nn.Linear(in_features=512, out_features=num_classes) self.relu = nn.ReLU()

def forward(self, x): # 1) 기본 블록과 풀링층 통과 x = self.b1(x) x = self.pool(x) x = self.b2(x) x = self.pool(x) x = self.b3(x) x = self.pool(x) # 2) 분류기의 입력으로 사용하기 위한 평탄화 x = torch.flatten(x, start_dim=1) # 3) 분류기로 예측값 출력 x = self.fc1(x) x = self.relu(x) x = self.fc2(x) x = self.relu(x) x = self.fc3(x) return x

import tqdm from torch.optim.adam import Adam from torch.utils.data.dataloader import DataLoader from torchvision.datasets.cifar import CIFAR10 from torchvision.transforms import ( Compose, Normalize, RandomCrop, RandomHorizontalFlip, ToTensor, ) transforms = Compose( [ RandomCrop((32, 32), padding=4), # 랜덤 크롭 RandomHorizontalFlip(p=0.5), # 랜덤 y축 대칭(수평 뒤집기) ToTensor(), Normalize( mean=(0.4914, 0.4822, 0.4465), std=(0.247, 0.243, 0.261), ), ] )

# 데이터셋 정의 training_data = CIFAR10( root="./", train=True, download=True, transform=transforms, ) test_data = CIFAR10( root="./", train=False, download=True, transform=transforms, ) # 데이터 로더 정의 train_loader = DataLoader(training_data, batch_size=32, shuffle=True) test_loader = DataLoader(test_data, batch_size=32, shuffle=False)

device = "cuda" if torch.cuda.is_available() else "cpu" model = ResNet(num_classes=10) model.to(device)

lr = 1e-4 optim = Adam(model.parameters(), lr=lr) for epoch in range(30): iterator = tqdm.tqdm(train_loader) for data, label in iterator: # 1) 최적화를 위해 기울기를 초기화 optim.zero_grad() # 2) 모델의 예측값 preds = model(data.to(device)) # 3) 손실 계산 및 역전파 loss = nn.CrossEntropyLoss()(preds, label.to(device)) loss.backward() optim.step() iterator.set_description(f"epoch:{epoch+1} loss:{loss.item()}") # 학습된 모델 저장 torch.save(model.state_dict(), "ResNet.pth")

model.load_state_dict(torch.load("ResNet.pth", map_location=device)) num_corr = 0 with torch.no_grad(): for data, label in test_loader: output = model(data.to(device)) preds = output.data.max(1)[1] corr = preds.eq(label.to(device).data).sum().item() num_corr += corr print(f"Accuracy: {num_corr / len(test_data)}")