[OpenCV] torch와 torchvision 라이브러리를 이용하여 Fashion MNIST에서의 Classificationtorch와 torchvision 라이브러리를 이용하여 Fashion MNIST에서의 Classification

작성자	김명원
일 시	2024. 5. 23  (목) 18:00 ~ 21:00
장 소	복지관 b128-1호
참가자 명단	임혜진, 이재영, 성창민, 김명원, 장원준
사 진

사용할 라이브러리 호출

import torch

import torchvision

import torchvision.transforms as transforms

import torch.nn as nn

import torch.nn.functional as F

import torch.optim as optim

데이터셋 로드 및 전처리

# 데이터셋 로드 및 전처리

transform = transforms.Compose([

transforms.ToTensor(),

transforms.Normalize((0.5,), (0.5,))

])

from torch.utils.data import DataLoader

trainset = torchvision.datasets.FashionMNIST(

root='./data',

train=True,

download=True,

transform=transform

)

trainloader = DataLoader(

trainset,

batch_size=4,

shuffle=True,

num_workers=2

)

testset = torchvision.datasets.FashionMNIST(

root='./data',

train=False,

download=True,

transform=transform

)

testloader = DataLoader(

testset,

batch_size=4,

shuffle=False,

num_workers=2

)

classes = ('T-shirt/top', 'Trouser', 'Pullover', 'Dress', 'Coat', 'Sandal', 'Shirt', 'Sneaker', 'Bag', 'Ankle boot')

신경망 정의

# 신경망 정의

class Net(nn.Module):

def __init__(self):

super(Net, self).__init__()

self.conv1 = nn.Conv2d(1, 6, 5)

self.pool = nn.MaxPool2d(2, 2)

self.conv2 = nn.Conv2d(6, 16, 5)

self.fc1 = nn.Linear(16 * 4 * 4, 120)

self.fc2 = nn.Linear(120, 84)

self.fc3 = nn.Linear(84, 10) #'hint: ~(~,class_num)'

def forward(self, x):

x = self.pool(F.relu(self.conv1(x)))

x = self.pool(F.relu(self.conv2(x))) #'hint: conv2'

x = x.view(-1, 16 * 4 * 4)

x = F.relu(self.fc1(x))

x = F.relu(self.fc2(x)) #'hint: fc2'

x = self.fc3(x)

return x

Loss, optimizer 정의

# 신경망 생성 및 손실 함수, 최적화 함수 정의

net = Net()

# pos_weight = torch.tensor([1.0])

# criterion = nn.BCEWithLogitsLoss(pos_weight=pos_weight), 'hint: crossentropy'

criterion = nn.CrossEntropyLoss()

# model = nn.Linear(in_features=1, out_features=1, bias=False)

optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)

모델 학습

# 2 epoch만큼 반복

for epoch in range(2):

running_loss = 0.0

for i, data in enumerate(trainloader, 0):

inputs, labels = data

# Gradient 매개변수를 0으로 만들고

optimizer.zero_grad()

# 순전파 + 역전파 + 최적화

outputs = net(inputs)

loss = criterion(outputs, labels)

loss.backward()

optimizer.step()

# 통계 출력

running_loss += loss.item()

if i % 200 == 199:

print(f'[{epoch + 1}, {i + 1:3d}] loss: {running_loss / 200:.3f}')

running_loss = 0.0

[1, 200] loss: 2.298 [1, 400] loss: 2.286 [1, 600] loss: 2.209 [1, 800] loss: 1.594 [1, 1000] loss: 1.080 [1, 1200] loss: 1.052 [1, 1400] loss: 0.866 [1, 1600] loss: 0.820 [1, 1800] loss: 0.724 [1, 2000] loss: 0.766 [1, 2200] loss: 0.767 [1, 2400] loss: 0.757 [1, 2600] loss: 0.722 [1, 2800] loss: 0.675 [1, 3000] loss: 0.702 [1, 3200] loss: 0.634 [1, 3400] loss: 0.669 [1, 3600] loss: 0.656 [1, 3800] loss: 0.653 [1, 4000] loss: 0.654 [1, 4200] loss: 0.620 [1, 4400] loss: 0.650 [1, 4600] loss: 0.550 [1, 4800] loss: 0.630 [1, 5000] loss: 0.620 [1, 5200] loss: 0.639 [1, 5400] loss: 0.578 [1, 5600] loss: 0.534 [1, 5800] loss: 0.579 [1, 6000] loss: 0.617 [1, 6200] loss: 0.569 [1, 6400] loss: 0.525 [1, 6600] loss: 0.547 [1, 6800] loss: 0.545 [1, 7000] loss: 0.531 [1, 7200] loss: 0.541 [1, 7400] loss: 0.551 [1, 7600] loss: 0.480 [1, 7800] loss: 0.466 [1, 8000] loss: 0.518 [1, 8200] loss: 0.555 [1, 8400] loss: 0.481 [1, 8600] loss: 0.515 [1, 8800] loss: 0.445 [1, 9000] loss: 0.530 [1, 9200] loss: 0.507 [1, 9400] loss: 0.451 [1, 9600] loss: 0.439 [1, 9800] loss: 0.468 [1, 10000] loss: 0.508 [1, 10200] loss: 0.486 [1, 10400] loss: 0.452 [1, 10600] loss: 0.485 [1, 10800] loss: 0.449 [1, 11000] loss: 0.463 [1, 11200] loss: 0.450 [1, 11400] loss: 0.509 [1, 11600] loss: 0.542 [1, 11800] loss: 0.470 [1, 12000] loss: 0.502 [1, 12200] loss: 0.493 [1, 12400] loss: 0.469 [1, 12600] loss: 0.453 [1, 12800] loss: 0.500 [1, 13000] loss: 0.404 [1, 13200] loss: 0.435 [1, 13400] loss: 0.441 [1, 13600] loss: 0.468 [1, 13800] loss: 0.453 [1, 14000] loss: 0.477 [1, 14200] loss: 0.403 [1, 14400] loss: 0.443 [1, 14600] loss: 0.386 [1, 14800] loss: 0.453 [1, 15000] loss: 0.428 [2, 200] loss: 0.381 [2, 400] loss: 0.448 [2, 600] loss: 0.417 [2, 800] loss: 0.377 [2, 1000] loss: 0.402 [2, 1200] loss: 0.398 [2, 1400] loss: 0.364 [2, 1600] loss: 0.380 [2, 1800] loss: 0.383 [2, 2000] loss: 0.365 [2, 2200] loss: 0.411 [2, 2400] loss: 0.413 [2, 2600] loss: 0.386 [2, 2800] loss: 0.395 [2, 3000] loss: 0.402 [2, 3200] loss: 0.396 [2, 3400] loss: 0.387 [2, 3600] loss: 0.389 [2, 3800] loss: 0.407 [2, 4000] loss: 0.369 [2, 4200] loss: 0.360 [2, 4400] loss: 0.405 [2, 4600] loss: 0.339 [2, 4800] loss: 0.356 [2, 5000] loss: 0.365 [2, 5200] loss: 0.335 [2, 5400] loss: 0.331 [2, 5600] loss: 0.401 [2, 5800] loss: 0.398 [2, 6000] loss: 0.396 [2, 6200] loss: 0.437 [2, 6400] loss: 0.414 [2, 6600] loss: 0.391 [2, 6800] loss: 0.378 [2, 7000] loss: 0.354 [2, 7200] loss: 0.378 [2, 7400] loss: 0.389 [2, 7600] loss: 0.361 [2, 7800] loss: 0.421 [2, 8000] loss: 0.341 [2, 8200] loss: 0.355 [2, 8400] loss: 0.394 [2, 8600] loss: 0.356 [2, 8800] loss: 0.401 [2, 9000] loss: 0.371 [2, 9200] loss: 0.360 [2, 9400] loss: 0.373 [2, 9600] loss: 0.355 [2, 9800] loss: 0.384 [2, 10000] loss: 0.347 [2, 10200] loss: 0.366 [2, 10400] loss: 0.415 [2, 10600] loss: 0.415 [2, 10800] loss: 0.410 [2, 11000] loss: 0.325 [2, 11200] loss: 0.344 [2, 11400] loss: 0.341 [2, 11600] loss: 0.358 [2, 11800] loss: 0.395 [2, 12000] loss: 0.351 [2, 12200] loss: 0.360 [2, 12400] loss: 0.316 [2, 12600] loss: 0.365 [2, 12800] loss: 0.324 [2, 13000] loss: 0.337 [2, 13200] loss: 0.327 [2, 13400] loss: 0.311 [2, 13600] loss: 0.343 [2, 13800] loss: 0.337 [2, 14000] loss: 0.354 [2, 14200] loss: 0.359 [2, 14400] loss: 0.323 [2, 14600] loss: 0.331 [2, 14800] loss: 0.352 [2, 15000] loss: 0.363

모델 저장 및 평가

# 모델 저장

PATH = './fashion_net.pth'

torch.save(net.state_dict(), PATH)

# 테스트 데이터로 모델 평가

correct = 0

total = 0

with torch.no_grad():

for data in testloader:

# 입력 받기

images, labels = data

outputs = net(images)

# 최대값 출력

_, predicted = torch.max(outputs, 1)

total += labels.size(0)

correct += (predicted == labels).sum().item()

print('Accuracy of the network on the 10000 test images: %d %%' % (100 * correct / total))

import matplotlib.pyplot as plt

import numpy as np

# 시각화

def imshow(img):

img = img / 2 + 0.5

npimg = img.numpy()

plt.imshow(np.transpose(npimg, (1, 2, 0)))

plt.show()

dataiter = iter(testloader)

for i in range(4):

images, labels = next(dataiter)

imshow(torchvision.utils.make_grid(images))

print('GroundTruth: ', ' '.join('%5s' % classes[labels[j]] for j in range(4)))

outputs = net(images)

_, predicted = torch.max(outputs, 1)

print('Predicted: ', ' '.join('%5s' % classes[predicted[j]] for j in range(4)))

드디어 마지막 모각코이다.
내가 공부하고 실습한 내용을 정리하면서
많은 공부가 됐고 앞으로 다시 필요한 내용을 찾을 때 좋을 것 같다.