VGGNet Implementation¶
- ÄÁº¼·ç¼Ç ¿¬»ê¸¸ ¹è¿î »óÅ¿¡¼ VGG¸¦ ¹Ù·Î ÀÌÇØÇÏ°í Â¥±â¿¡´Â ¹«¸®°¡ ÀÖ½À´Ï´Ù.
¿¬»êµéÀÇ µ¿ÀÛ ¿ø¸®¸¦ ÃæºÐÈ÷ ÀÌÇØÇÑÈÄ ´Ù½Ã º¸¼Åµµ ´ÊÁö ¾Ê½À´Ï´Ù.
2014 ILSVRC 2nd place
- VGG-16
- Convolution layer
- Maxpooling layer
- Fully connected layer
In [1]:
# ·±Å¸ÀÓ À¯Çü GPU ¸ðµå·Î º¯°æ
!pip install torch torchvision
Prepare Data¶
- ¸ðµ¨ÀÌ ÇнÀÀÌ µÇ´ÂÁö¸¸ È®ÀÎÇÒ ¼ö ÀÖ°Ô °£´ÜÇÑ µ¥ÀÌÅ͸¦ ´Ù¿î·Îµå ÇÕ´Ï´Ù.
In [2]:
!rm -r images
import os
# À̹ÌÁö ÆÄÀÏÀ» ÀúÀåÇÒ Æú´õ¸¦ »ý¼ºÇÕ´Ï´Ù.
try:
os.mkdir("images")
os.mkdir("images/dogs")
os.mkdir("images/cats")
except:
pass
# À̹ÌÁöµéÀ» ÁöÁ¤ÇÑ À§Ä¡¿¡ ´Ù¿î·ÎµåÇÕ´Ï´Ù.
# images/dogs ¹Ø¿¡ 2°³
!wget https://i.kinja-img.com/gawker-media/image/upload/s--WFkXeene--/c_scale,f_auto,fl_progressive,q_80,w_800/ol9ceoqxidudap8owlwn.jpg -P images/dogs
!wget https://www.rspcansw.org.au/wp-content/uploads/2017/08/50_a-feature_dogs-and-puppies_mobile.jpg -P images/dogs
# images/cats ¹Ø¿¡ 2°³
!wget https://www.catster.com/wp-content/uploads/2018/05/A-gray-cat-crying-looking-upset.jpg -P images/cats
!wget https://www.scarymommy.com/wp-content/uploads/2018/01/c1.jpg?w=700 -P images/cats
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 torch.utils.data as data
import torchvision.datasets as dset
import torchvision.transforms as transforms
from torch.utils.data import DataLoader
2) Hyperparameter¶
In [0]:
batch_size= 1
learning_rate = 0.0002
num_epoch = 100
2. Data Loader¶
- https://pytorch.org/docs/stable/torchvision/datasets.html?highlight=imagefolder#torchvision.datasets.ImageFolder
- ImageFolder¶ó´Â ÇÔ¼ö¸¦ ÀÌ¿ëÇØ µû·Î À̹ÌÁö-¶óº§ ½ÖÀ» ¸¸µéÁö ¾Ê°í Æú´õ¿¡ ÀúÀåÇϴ°͸¸À¸·Î ½±°Ô À̹ÌÁö-¶óº§ ½ÖÀ» ¸¸µé ¼ö ÀÖ½À´Ï´Ù.
ex)
root/dog/xxx.png
root/dog/xxy.png
root/cat/123.png
root/cat/nsdf3.png
In [0]:
# ¶óº§(ȤÀº Ŭ·¡½º) º°·Î Æú´õ°¡ ÀúÀåµÇ¾î ÀÖ´Â ·çÆ® µð·ºÅ丮¸¦ ÁöÁ¤ÇÕ´Ï´Ù.
img_dir = "./images"
# ÇØ´ç ·çÆ® µð·ºÅ丮¸¦ ImageFolder ÇÔ¼ö¿¡ Àü´ÞÇÕ´Ï´Ù.
# À̶§ À̹ÌÁöµé¿¡ ´ëÇÑ º¯Çüµµ °°ÀÌ Àü´ÞÇØÁÝ´Ï´Ù.
img_data = dset.ImageFolder(img_dir, transforms.Compose([
transforms.Resize(256), # À̹ÌÁö Å©±â¸¦ 256x256À¸·Î ¹Ù²ãÁÝ´Ï´Ù.
transforms.RandomResizedCrop(224), # 256x256 À̹ÌÁöÀÇ ·£´ýÇÑ À§Ä¡¿¡¼ 224x224 Å©±â¸¸Å »ùÇøµ ÇÕ´Ï´Ù.
transforms.RandomHorizontalFlip(), # ·£´ýÇÑ È®·ü·Î À̹ÌÁö¸¦ Á¿ì¹ÝÀü ÇÕ´Ï´Ù.
transforms.ToTensor(), # À̹ÌÁö µ¥ÀÌÅ͸¦ ÅÙ¼·Î º¯ÇüÇÕ´Ï´Ù.
]))
train_loader = data.DataLoader(img_data, batch_size=batch_size,
shuffle=True, num_workers=2)
In [0]:
# ÄÁº¼·ç¼Ç ¿¬»êÀÌ 2¹ø ¿¬¼ÓÇÏ´Â °æ¿ì
# ÄÁº¼·ç¼Ç-È°¼ºÈÇÔ¼ö-ÄÁº¼·ç¼Ç-È°¼ºÈÇÔ¼ö-Ç®¸µ
def conv_2_block(in_dim,out_dim):
model = nn.Sequential(
nn.Conv2d(in_dim,out_dim,kernel_size=3,padding=1),
nn.ReLU(),
nn.Conv2d(out_dim,out_dim,kernel_size=3,padding=1),
nn.ReLU(),
nn.MaxPool2d(2,2)
)
return model
# ÄÁº¼·ç¼Ç ¿¬»êÀÌ 3¹ø ¿¬¼ÓÇÏ´Â °æ¿ì
# ÄÁº¼·ç¼Ç-È°¼ºÈÇÔ¼ö-ÄÁº¼·ç¼Ç-È°¼ºÈÇÔ¼ö-ÄÁº¼·ç¼Ç-È°¼ºÈÇÔ¼ö-Ç®¸µ
def conv_3_block(in_dim,out_dim):
model = nn.Sequential(
nn.Conv2d(in_dim,out_dim,kernel_size=3,padding=1),
nn.ReLU(),
nn.Conv2d(out_dim,out_dim,kernel_size=3,padding=1),
nn.ReLU(),
nn.Conv2d(out_dim,out_dim,kernel_size=3,padding=1),
nn.ReLU(),
nn.MaxPool2d(2,2)
)
return model
2) VGG Model¶
In [0]:
# À§¿¡¼ Á¤ÀÇÇÑ ºí·ÏµéÀ» ÀÌ¿ëÇØ VGG ³×Æ®¿öÅ©¸¦ ¸¸µé¾îº¸°Ú½À´Ï´Ù.
# ÇÊÅÍÀÇ °³¼ö°¡ 2ÀÇ n½ÂÀÇ °ªÀ» °¡Áö±â ¶§¹®¿¡ base_dimÀ̶õ º¯¼ö¸¦ Ãß°¡Çؼ ´Ü¼øÈ Çß½À´Ï´Ù.
# ÇöÀç dog, cat µÎ °¡Áö Ŭ·¡½º¸¦ ±¸ºÐÇÏ·Á°í Çϱ⠶§¹®¿¡ num_classes=2·Î ¼³Á¤Çß½À´Ï´Ù.
class VGG(nn.Module):
def __init__(self, base_dim, num_classes=2):
super(VGG, self).__init__()
self.feature = nn.Sequential(
conv_2_block(3,base_dim),
conv_2_block(base_dim,2*base_dim),
conv_3_block(2*base_dim,4*base_dim),
conv_3_block(4*base_dim,8*base_dim),
conv_3_block(8*base_dim,8*base_dim),
)
self.fc_layer = nn.Sequential(
nn.Linear(8*base_dim * 7 * 7, 100),
nn.ReLU(True), # True ´Â inplace ¿¬»êÀ» ÇÏ°Ú´Ù´Â Àǹ̸¦ °¡Áý´Ï´Ù. inplace ¿¬»êÀº °á°ú°ªÀ» »õ·Î¿î º¯¼ö¿¡ °ªÀ» ÀúÀåÇÏ´Â ´ë½Å ±âÁ¸ÀÇ µ¥ÀÌÅ͸¦ ´ëüÇϴ°ÍÀ» ÀǹÌÇÕ´Ï´Ù.
#nn.Dropout(),
nn.Linear(100, 20),
nn.ReLU(True),
#nn.Dropout(),
nn.Linear(20, num_classes),
)
def forward(self, x):
x = self.feature(x)
x = x.view(x.size(0), -1) # x.size(0)¸¦ batch size·Î ¹Ù²ãµµ °°Àº °ªÀÔ´Ï´Ù.
x = self.fc_layer(x)
return x
4. Optimizer & Loss¶
In [8]:
# gpu°¡ »ç¿ë °¡´ÉÇÑ °æ¿ì¿¡´Â device¸¦ 0¹ø gpu·Î ¼³Á¤ÇÏ°í ºÒ°¡´ÉÇϸé cpu·Î ¼³Á¤ÇÕ´Ï´Ù.
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print(device)
# ¾Õ¼ Á¤ÀÇÇÑ´ë·Î vGG Ŭ·¡½º¸¦ ÀνºÅϽºÈ ÇÏ°í ÁöÁ¤ÇÑ ÀåÄ¡¿¡ ¿Ã¸³´Ï´Ù.
model = VGG(base_dim=16).to(device)
# ¼Õ½ÇÇÔ¼ö ¹× ÃÖÀûÈÇÔ¼ö¸¦ ¼³Á¤ÇÕ´Ï´Ù.
loss_func = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
# ¸ðµ¨ ÀÚ³à ³ëµåÀÇ À̸§°ú ¸ðµâÀ» Ãâ·ÂÇÕ´Ï´Ù.
for i in model.named_children():
print(i)
5. 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)
In [0]: