정형화 Weight Regularization

  • ÃÖÀûÈ­ ÇÔ¼öÀÇ weight_decay ·Î °­µµ¸¦ Á¶ÀýÇÒ ¼ö ÀÖ½À´Ï´Ù.
  • ex) torch.optim.SGD(params, lr=1, momentum=0, dampening=0, weight_decay=0, nesterov=False)
  • ¸ðµ¨ÀÌ ¿À¹öÇÇÆÃÇÒ °æ¿ì, ÀûÀýÇÑ °­µµ·Î Á¤ÇüÈ­¸¦ °É¾îÁÖ¸é À̸¦ ¾î´ÀÁ¤µµ ±Øº¹ÇÒ ¼ö ÀÖ½À´Ï´Ù.
  • Á¤ÇüÈ­ ºÎºÐ »©°í´Â ÄÁº¼·ç¼Ç Àΰø½Å°æ¸Á ÄÚµå¿Í µ¿ÀÏÇÕ´Ï´Ù.
In [1]:
# ·±Å¸ÀÓ À¯ÇüÀ» GPU·Î ¹Ù²Ù½Ã±æ Ãßõµå¸³´Ï´Ù.
!pip install torch torchvision

Requirement already satisfied: torch in /usr/local/lib/python3.6/dist-packages (1.1.0)
Requirement already satisfied: torchvision in /usr/local/lib/python3.6/dist-packages (0.3.0)
Requirement already satisfied: numpy in /usr/local/lib/python3.6/dist-packages (from torch) (1.16.4)
Requirement already satisfied: pillow>=4.1.1 in /usr/local/lib/python3.6/dist-packages (from torchvision) (4.3.0)
Requirement already satisfied: six in /usr/local/lib/python3.6/dist-packages (from torchvision) (1.12.0)
Requirement already satisfied: olefile in /usr/local/lib/python3.6/dist-packages (from pillow>=4.1.1->torchvision) (0.46)

1. Settings

1) Import required libraries

In [0]:
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.init as init
import torchvision.datasets as dset
import torchvision.transforms as transforms
from torch.utils.data import DataLoader

2) Set hyperparameters

In [0]:
batch_size = 256
learning_rate = 0.0002
num_epoch = 10

2. Data

1) Download Data

In [0]:
mnist_train = dset.MNIST("./", train=True, transform=transforms.ToTensor(), target_transform=None, download=True)
mnist_test = dset.MNIST("./", train=False, transform=transforms.ToTensor(), target_transform=None, download=True)

2) Check Dataset

In [5]:
print(mnist_train.__getitem__(0)[0].size(), mnist_train.__len__())
mnist_test.__getitem__(0)[0].size(), mnist_test.__len__()

torch.Size([1, 28, 28]) 60000
Out[5]:
(torch.Size([1, 28, 28]), 10000)

3) Set DataLoader

In [0]:
train_loader = torch.utils.data.DataLoader(mnist_train,batch_size=batch_size, shuffle=True,num_workers=2,drop_last=True)
test_loader = torch.utils.data.DataLoader(mnist_test,batch_size=batch_size, shuffle=False,num_workers=2,drop_last=True)

3. Model & Optimizer

1) CNN Model

In [0]:
class CNN(nn.Module):
    def __init__(self):
        super(CNN,self).__init__()
        self.layer = nn.Sequential(
            nn.Conv2d(1,16,3,padding=1),  # 28 x 28
            nn.ReLU(),
            nn.Conv2d(16,32,3,padding=1), # 28 x 28
            nn.ReLU(),
            nn.MaxPool2d(2,2),            # 14 x 14
            nn.Conv2d(32,64,3,padding=1), # 14 x 14
            nn.ReLU(),
            nn.MaxPool2d(2,2)             #  7 x 7
        )
        self.fc_layer = nn.Sequential(
            nn.Linear(64*7*7,100),
            nn.ReLU(),
            nn.Linear(100,10)
        )       
        
    def forward(self,x):
        out = self.layer(x)
        out = out.view(batch_size,-1)
        out = self.fc_layer(out)
        return out

2) Loss func & Optimizer

In [8]:
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print(device)

model = CNN().to(device)
loss_func = nn.CrossEntropyLoss()

# Á¤ÇüÈ­´Â weight_decay·Î ÁÙ ¼ö ÀÖ½À´Ï´Ù.
optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate, weight_decay=0.1)

cuda:0

4. Train

In [9]:
for i in range(num_epoch):
    for j,[image,label] in enumerate(train_loader):
        x = image.to(device)
        y_= label.to(device)
        
        optimizer.zero_grad()
        output = model.forward(x)
        loss = loss_func(output,y_)
        loss.backward()
        optimizer.step()
        
    if i % 10 == 0:
        print(loss)          

tensor(2.3000, device='cuda:0', grad_fn=<NllLossBackward>)
In [0]:
#param_list = list(model.parameters())
#print(param_list)

5. Test

In [11]:
correct = 0
total = 0

with torch.no_grad():
  for image,label in test_loader:
      x = image.to(device)
      y_= label.to(device)

      output = model.forward(x)
      _,output_index = torch.max(output,1)

      total += label.size(0)
      correct += (output_index == y_).sum().float()

  print("Accuracy of Test Data: {}".format(100*correct/total))

Accuracy of Test Data: 9.805688858032227

정확도가 왜 이렇게 낮을까?

  • ¸ÕÀú ´Ù¸¥ ±â¹ýµéÀ» ¹è¿öº¸¸é ÀÌÀ¯¸¦ ¾Ë ¼ö ÀÖ½À´Ï´Ù!
In [0]: