PyTorch로 VGG network 구현하기
2022. 12. 7. 17:00ㆍ딥러닝/딥러닝 모델
In [2]:
# my github: https://github.com/withAnewWorld/models_from_scratch
# my blog
# https://self-deeplearning.blogspot.com/
# https://self-deeplearning.tistory.com/
import torch
import torch.nn as nn
import torch.nn.functional as F
Ref¶
paper: https://arxiv.org/pdf/1409.1556.pdf
youtube(Aladdin Persson):
https://www.youtube.com/watch?v=ACmuBbuXn20&list=PLhhyoLH6IjfxeoooqP9rhU3HJIAVAJ3Vz&index=17
목차¶
- VGG Network 의의
- VGG Network architecture
VGG Network¶
의의
1) Alex Net(conv net기반 model의 전성기를 열은 network)이후 conv net기반 model의 layer를 깊게 쌓을수록 성능이 좋아질 것이라는 예측을 실험적으로 보임
2) 작은 kernel_size (3x3)를 이용하여 model의 성능을 끌어올림
VGG network architecture¶
VGG architecture from VGG paper table 1
VGG net의 경우 여러 layer의 갯수에 따라 version이 있습니다. (11, 13, 16, 19)
모든 conv net은 kernel_size: 3x3, stride: 1, padding: 1로 고정입니다.
(-> Height, Width 고정)
다만, Max Pool layer를 거칠 때마다 Height와 Width가 각각 반으로 줄어듭니다.
(kernel_size: 2, stride: 2)
table 1.을 보시면 저자가 왼쪽에서부터 오른쪽으로 새로운 layer가 추가(or 변경)되는 것에 대해 굵은 글씨체로 써줬습니다.
따라서 version이 변경될 때마다 굵은 글씨의 layer를 변경하면 될 것 같습니다.
cf) VGG16의 conv1(kernel_size: 1)의 경우 model 비교를 위해 저자가 설정한 것으로 생략하겠습니다.
in spite of the same depth, the configuration C (which contains three 1 × 1 conv. layers), performs worse than the configuration D, (from paper 6page)
# pseudo code
class VGG(version, ...):
1) VGG_11을 설계한다.
2) if version > 11:
layer 추가
3) if version > 13:
layer 추가
4) if version > 16:
layer 추가
5) if version == 19:
layer 추가
def forward(self, x):
'''
input:
- x(Tensor[N, C, 224, 224])
output:
- logits(Tensor[N, num_classes])
'''
그러면 VGG 11을 pseudo code로 설계해보겠습니다.
# pseudo code
class VGG(nn.Module):
def __init__(self, version, ...):
super(VGG, self).__init__()
self.block_1 = nn.Conv2d() C(3->64), depth: 1
self.pool = nn.MaxPool2d() (Height, Width: 224 -> 112)
self.block_2 = nn.Conv2d() C(64->128), depth: 1
self.pool = MaxPool2d() (Height, Width: 112 -> 56)
self.block_3 = nn.Conv2d() C(128->256) depth: 2
self.pool = MaxPool2d() (Height, Width: 56 -> 28)
self.block_4 = nn.Conv2d() C(256->512) depth: 2
self.pool = MaxPool2d() (Height, Width: 28 -> 14)
self.block_5 = nn.Conv2d() C(512->512) depth: 2
self.pool = MaxPool2d() (Height, Width: 14 -> 7)
self.linear_block = nn.Sequential(
nn.Flatten(),
nn.Linear(in_features = 512*7*7(C*H*W), out_featuers = 4096),
nn.Linear(in_features = 4096, out_featuers = 4096),
nn.Linear(in_features = 4096, out_featuers = num_classes(1000)),
nn.Softmax(dim = -1)
)
def forward(self, x):
pass
In [ ]:
class VGG(nn.Module):
def __init__(self, version, in_channels, H, W):
'''
inputs:
- verison(int): one of (11, 13, 16, 19)
- in_channels: img channels (ex. RGB: 3)
- H: img height
- W: img width
'''
super().__init__()
num_pool = 5
self.block_1 = nn.ModuleList([
nn.Conv2d(in_channels = in_channels,
out_channels = 64,
kernel_size = (3, 3),
stride = 1,
padding = 1),
nn.MaxPool2d(kernel_size = (2, 2),
stride = 2)
])
if version > 11:
self.block_1.insert(1, nn.Conv2d(in_channels = 64,
out_channels = 64,
kernel_size = (3, 3),
stride = 1,
padding = 1))
self.block_2 = nn.ModuleList([
nn.Conv2d(in_channels = 64,
out_channels = 128,
kernel_size = (3,3),
stride = 1,
padding = 1),
nn.MaxPool2d(kernel_size = (2, 2),
stride = 2)
])
if version > 11:
self.block_2.insert(1, nn.Conv2d(in_channels = 128,
out_channels = 128,
kernel_size = (3,3),
stride = 1,
padding = 1))
self.block_3 = nn.ModuleList([
nn.Conv2d(in_channels = 128,
out_channels = 256,
kernel_size = (3,3),
stride = 1,
padding = 1),
nn.Conv2d(in_channels = 256,
out_channels = 256,
kernel_size = (3, 3),
stride = 1,
padding = 1),
nn.MaxPool2d(kernel_size = (2, 2),
stride = 2)
])
if version > 13:
self.block_3.insert(2, nn.Conv2d(in_channels = 256,
out_channels = 256,
kernel_size = (3, 3),
stride = 1,
padding = 1))
if version > 16:
self.block_3.insert(3, nn.Conv2d(in_channels = 256,
out_channels = 256,
kernel_size = (3, 3),
stride = 1,
padding = 1))
self.block_4 = nn.ModuleList([
nn.Conv2d(in_channels = 256,
out_channels = 512,
kernel_size = (3,3),
stride = 1,
padding = 1),
nn.Conv2d(in_channels = 512,
out_channels = 512,
kernel_size = (3,3),
stride = 1,
padding = 1),
nn.MaxPool2d(kernel_size = (2, 2),
stride = 2)
])
if version > 13:
self.block_4.insert(2, nn.Conv2d(in_channels = 512,
out_channels = 512,
kernel_size = (3,3),
stride = 1,
padding = 1))
if version > 16:
self.block_4.insert(3, nn.Conv2d(in_channels = 512,
out_channels = 512,
kernel_size = (3,3),
stride = 1,
padding = 1))
self.block_5 = nn.ModuleList([
nn.Conv2d(in_channels = 512,
out_channels = 512,
kernel_size = (3, 3),
stride = 1,
padding = 1),
nn.Conv2d(in_channels = 512,
out_channels = 512,
kernel_size = (3, 3),
stride = 1,
padding = 1),
nn.MaxPool2d(kernel_size = (2, 2),
stride = 2)
])
if version > 13:
self.block_5.insert(2, nn.Conv2d(in_channels = 512,
out_channels = 512,
kernel_size = (3, 3),
stride = 1,
padding = 1))
if version > 16:
self.block_5.insert(2, nn.Conv2d(in_channels = 512,
out_channels = 512,
kernel_size = (3, 3),
stride = 1,
padding = 1))
self.linear_block = nn.Sequential(
nn.Flatten(),
nn.Linear(in_features = 512*H//(2**(num_pool))*W//(2**(num_pool)),
out_features = 4096),
nn.Linear(in_features = 4096,
out_features = 4096),
nn.Linear(in_features = 4096,
out_features = 1000)
)
def forward(self, x):
for module in self.block_1:
x = module(x)
print(x.size())
for module in self.block_2:
x = module(x)
print(x.size())
for module in self.block_3:
x = module(x)
print(x.size())
for module in self.block_4:
x = module(x)
print(x.size())
for module in self.block_5:
x = module(x)
print(x.size())
x = self.linear_block(x)
print(x.size())
In [ ]:
x = torch.randn(1, 3, 224, 224)
model = VGG(19, x.size(1), x.size(2), x.size(3))
In [ ]:
model
Out[ ]:
VGG(
(block_1): ModuleList(
(0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(2): MaxPool2d(kernel_size=(2, 2), stride=2, padding=0, dilation=1, ceil_mode=False)
)
(block_2): ModuleList(
(0): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(2): MaxPool2d(kernel_size=(2, 2), stride=2, padding=0, dilation=1, ceil_mode=False)
)
(block_3): ModuleList(
(0): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(4): MaxPool2d(kernel_size=(2, 2), stride=2, padding=0, dilation=1, ceil_mode=False)
)
(block_4): ModuleList(
(0): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(4): MaxPool2d(kernel_size=(2, 2), stride=2, padding=0, dilation=1, ceil_mode=False)
)
(block_5): ModuleList(
(0): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(4): MaxPool2d(kernel_size=(2, 2), stride=2, padding=0, dilation=1, ceil_mode=False)
)
(linear_block): Sequential(
(0): Flatten(start_dim=1, end_dim=-1)
(1): Linear(in_features=25088, out_features=4096, bias=True)
(2): Linear(in_features=4096, out_features=4096, bias=True)
(3): Linear(in_features=4096, out_features=1000, bias=True)
)
)In [ ]:
model(x)
torch.Size([1, 64, 112, 112])
torch.Size([1, 128, 56, 56])
torch.Size([1, 256, 28, 28])
torch.Size([1, 512, 14, 14])
torch.Size([1, 512, 7, 7])
torch.Size([1, 1000])
'딥러닝 > 딥러닝 모델' 카테고리의 다른 글
| PyTorch로 Transformer 구현하기 (0) | 2022.12.07 |
|---|---|
| PyTorch로 EfficientNet 구현하기 (0) | 2022.12.07 |
| PyTorch로 ResNet 구현하기 (0) | 2022.12.07 |
| PyTorch로 GoogLeNet 구현하기 (0) | 2022.12.07 |
| PyTorch로 딥러닝 모델 구현하는 방법(LeNet) (0) | 2022.12.07 |