Cnn 6 (cat and dog (2) vggnet)
Cat and Dog Classifier
from IPython.display import Image
Image('./images/dog_cat.png', width=600)
 VGGNet)/output_1_0.png)
import torch # 파이토치 기본 라이브러리
# torchvision : 데이터셋, 모델 아키텍처, 컴퓨터 비전의 이미지 변환 기능 제공
from torchvision import datasets # torchvision에서 제공하는 데이터셋
from torchvision import transforms # 이미지 변환기능을 제공하는 패키지
# torch.utils.data : 파이토치 데이터 로딩 유틸리티
from torch.utils.data import DataLoader # 모델 훈련에 사용할 수 있는 미니 배치 구성하고
# 매 epoch마다 데이터를 샘플링, 병렬처리 등의 일을 해주는 함수
from torch.utils.data import random_split
import numpy as np
import matplotlib.pyplot as plt
from torch.utils.tensorboard import SummaryWriter
from copy import deepcopy
!nvidia-smi
Tue Apr 11 02:54:06 2023
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 525.85.12 Driver Version: 525.85.12 CUDA Version: 12.0 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 Tesla T4 Off | 00000000:00:04.0 Off | 0 |
| N/A 34C P8 9W / 70W | 0MiB / 15360MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
| No running processes found |
+-----------------------------------------------------------------------------+
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
device
device(type='cuda')
0. 데이터 다운로드
# option 1
# from google.colab import files
# files.upload()
%pwd
'/content'
# option 2
# from google.colab import drive
# drive.mount('/content/drive')
# !cp '/content/drive/MyDrive/Classroom/Playdata 인공지능 28기/CNN/kaggle_catanddog.zip' './'
# !unzip -q kaggle_catanddog.zip -d catanddog/
# option 3
# kaggle api를 사용할 수 있는 패키지 설치
!pip install kaggle
# kaggle.json upload
from google.colab import files
files.upload()
# permmision warning 방지
!mkdir -p ~/.kaggle
!cp kaggle.json ~/.kaggle/
!chmod 600 ~/.kaggle/kaggle.json
# download
!kaggle datasets download -d tongpython/cat-and-dog
# unzip(압축풀기)
!unzip -q cat-and-dog.zip -d catanddog/
Saving kaggle.json to kaggle.json
Downloading cat-and-dog.zip to /content
97% 211M/218M [00:01<00:00, 215MB/s]
100% 218M/218M [00:01<00:00, 214MB/s]
data_dir = './catanddog/'
1. 데이터 불러오기
# Compose를 통해 원하는 전처리기를 차례대로 넣을 수 있음
transform = transforms.Compose([transforms.Resize([224, 224]), transforms.ToTensor()])
trainset = datasets.ImageFolder(data_dir+ 'training_set/training_set', transform = transform)
testset = datasets.ImageFolder(data_dir+ 'test_set/test_set', transform = transform)
print(type(trainset), len(trainset))
print(type(testset), len(testset))
<class 'torchvision.datasets.folder.ImageFolder'> 8005
<class 'torchvision.datasets.folder.ImageFolder'> 2023
print(type(trainset.targets), len(trainset.targets), trainset.targets[:5], trainset.targets[-5:])
<class 'list'> 8005 [0, 0, 0, 0, 0] [1, 1, 1, 1, 1]
# 클래스별 분포
for i in range(2): # 클래스별 순회
print('클래스(레이블)별 데이터 개수 : ', i, (np.array(trainset.targets) == i).sum())
클래스(레이블)별 데이터 개수 : 0 4000
클래스(레이블)별 데이터 개수 : 1 4005
from sklearn.model_selection import train_test_split
train_indices, valid_indices, _, _ = train_test_split(
range(len(trainset)), # X의 index
trainset.targets, # y
stratify=trainset.targets, # target의 비율이 train과 valid에 그대로 반영되게
test_size= 0.2, random_state=42)
len(train_indices), len(valid_indices) # 80%, 20%
(6404, 1601)
from torch.utils.data import Subset
train_set = Subset(trainset, train_indices)
valid_set = Subset(trainset, valid_indices)
valid_set[0][1] # 0번째 샘플의 정답
1
# 클래스별 분포
class_list = []
for i in range(2): # 클래스별 순회
s = 0
for j in range(len(valid_set)): # valid_data 1601개 순회
if valid_set[j][1] == i :
s += 1
class_list.append(s)
class_list
[800, 801]
# trainset을 다시 train용과 valid 용으로 나누고자 할 때
# trainset, validset = random_split(trainset, [50000, 10000])
print(type(train_set), len(train_set))
print(type(valid_set), len(valid_set))
print(type(testset), len(testset))
<class 'torch.utils.data.dataset.Subset'> 6404
<class 'torch.utils.data.dataset.Subset'> 1601
<class 'torchvision.datasets.folder.ImageFolder'> 2023
# 0번째 샘플에 2개의 원소가 있는데, 그중 첫번째 원소는 이미지, 두번째 원소는 정답
# 그러나 파이토치로 읽어들인 이미지 텐서의 형상이 channels * height * width 임
# 그에 비해 opencv, matplotlib으로 읽어들인 이미지 array의 형상은 height * width * channels
print(train_set[0][0].size(), train_set[0][1])
torch.Size([3, 224, 224]) 0
2. 데이터 시각화
labels_map = {0 : 'cat', 1 : 'dog'} # for cat and dog
figure, axes = plt.subplots(nrows=4, ncols=8, figsize=(14, 8))
axes = axes.flatten()
for i in range(32):
rand_i = np.random.randint(0, len(trainset))
image, label= trainset[rand_i][0].permute(1, 2, 0), trainset[rand_i][1]
axes[i].axis('off')
axes[i].imshow(image)
axes[i].set_title(labels_map[label])
 VGGNet)/output_26_0.png)
3. 데이터 적재
DataLoader
- 모델 훈련에 사용할 수 있는 미니 배치 구성하고
- 매 epoch마다 데이터를 샘플링, 병렬처리 등의 일을 해주는 함수
배치 사이즈
- 배치 사이즈 중요한 하이퍼 파라미터
-
16 이하로 사용하는것이 성능에 좋다고 알려져 있음
- 배치 사이즈가 크다는 것은 실제 Loss, Gradient, Weight를 구하는 데 참여하는 데이타가 많다라는 뜻
-
배치 사이즈가 작을 수록 모델이 학습을 하는데 한번도 보지 않은 신선한 데이터가 제공될 확률이 큼
- 배치 사이즈가 크면 학습시간은 줄일 수 있으나 적절한 배치사이즈로 학습을 해야 성능을 높일 수 있음
- (60000개의 데이터를 100개의 미니배치로 학습하면 1 epoch당 걸리는 횟수가 600번인데, 10개의 미니배치로 학습하면 1 epoch당 걸리는 횟수가 6000번)
batch_size = 16 # 100 -> 16
# dataloader = DataLoader(데이터셋, 배치사이즈, 셔플여부.....)
trainloader = DataLoader(train_set, batch_size=batch_size, shuffle=True) # 훈련용 50000개의 데이터를 100개씩 준비
validloader = DataLoader(valid_set, batch_size=batch_size, shuffle=False) # 검증용 10000개의 데이터를 100개씩 준비
testloader = DataLoader(testset, batch_size=batch_size, shuffle=False) # 테스트용 10000개의 데이터를 100개씩 준비
6404/16, 1601/16, 2023/16
(400.25, 100.0625, 126.4375)
print(type(trainloader), len(trainloader))
print(type(validloader), len(validloader))
print(type(testloader), len(testloader))
<class 'torch.utils.data.dataloader.DataLoader'> 401
<class 'torch.utils.data.dataloader.DataLoader'> 101
<class 'torch.utils.data.dataloader.DataLoader'> 127
train_iter = iter(trainloader)
images, labels = next(train_iter)
images.size(), labels.size()
(torch.Size([16, 3, 224, 224]), torch.Size([16]))
4. 모델 생성
from IPython.display import Image
Image('./images/VGG1.jpg', width=700)
 VGGNet)/output_35_0.jpg)
Image('./images/VGG5.jpg')
 VGGNet)/output_36_0.jpg)
Image('./images/vgg16.png', width=350)
 VGGNet)/output_37_0.png)
import torch.nn as nn # 파이토치에서 제공하는 다양한 계층 (Linear Layer, ....)
import torch.optim as optim # 옵티마이저 (경사하강법...)
import torch.nn.functional as F # 파이토치에서 제공하는 함수(활성화 함수...)
# 가중치 초기화
# https://pytorch.org/docs/stable/nn.init.html
# 현재 default 값
# Linear :
# https://github.com/pytorch/pytorch/blob/9cf62a4b5d3b287442e70c0c560a8e21d8c3b189/torch/nn/modules/linear.py#L168
# Conv :
# https://github.com/pytorch/pytorch/blob/9cf62a4b5d3b287442e70c0c560a8e21d8c3b189/torch/nn/modules/conv.py#L111
# 가중치 초기화시 고려할 사항
# 1. 값이 충분히 작아야 함
# 2. 값이 하나로 치우쳐선 안됨
# 3. 적당한 분산으로 골고루 분포가 되어야 함
완전 연결망과 CNN망과의 차이점
- 지역 연산
- 가중치 공유(적은 파라미터)
- 평행 이동 불변성
변경 사항
- 기존 AlexNet 대비 Feature map이 작아지고 모델 파라미터가 9800만개에서 4200만개 정도로 줄었음
- AlexNet에서는 82~83% 정확도
- VGG16에서는 83~84% 정확도
conv block 별 사이즈 확인
conv_block1 = nn.Sequential(
nn.Conv2d(in_channels=3, out_channels=64, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=64),
nn.ReLU(),
nn.Conv2d(in_channels=64, out_channels=64, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=64),
nn.ReLU(),
nn.MaxPool2d(kernel_size=2, stride=2)
) # [16, 64, 112, 112]
conv_block1_out = conv_block1(images)
conv_block1_out.shape
torch.Size([16, 64, 112, 112])
conv_block2 = nn.Sequential(
nn.Conv2d(in_channels=64, out_channels=128, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=128),
nn.ReLU(),
nn.Conv2d(in_channels=128, out_channels=128, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=128),
nn.ReLU(),
nn.MaxPool2d(kernel_size=2, stride=2)
) # [16, 128, 56, 56]
conv_block2_out = conv_block2(conv_block1_out)
conv_block2_out.shape
torch.Size([16, 128, 56, 56])
conv_block3 = nn.Sequential(
nn.Conv2d(in_channels=128, out_channels=256, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=256),
nn.ReLU(),
nn.Conv2d(in_channels=256, out_channels=256, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=256),
nn.ReLU(),
nn.Conv2d(in_channels=256, out_channels=512, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=512),
nn.ReLU(),
nn.MaxPool2d(kernel_size=2, stride=2)
) # [16, 512, 28, 28]
conv_block3_out = conv_block3(conv_block2_out)
conv_block3_out.shape
torch.Size([16, 512, 28, 28])
conv_block4 = nn.Sequential(
nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=512),
nn.ReLU(),
nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=512),
nn.ReLU(),
nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=512),
nn.ReLU(),
nn.MaxPool2d(kernel_size=2, stride=2)
) # [16, 512, 14, 14]
conv_block4_out = conv_block4(conv_block3_out)
conv_block4_out.shape
torch.Size([16, 512, 14, 14])
conv_block5 = nn.Sequential(
nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=512),
nn.ReLU(),
nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=512),
nn.ReLU(),
nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=512),
nn.ReLU(),
nn.MaxPool2d(kernel_size=2, stride=2)
) # [16, 512, 7, 7]
conv_block5_out = conv_block5(conv_block4_out)
conv_block5_out.shape
torch.Size([16, 512, 7, 7])
class VGG16(nn.Module):
def __init__(self):
super().__init__()
self.conv_block1 = nn.Sequential(
nn.Conv2d(in_channels=3, out_channels=64, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=64),
nn.ReLU(),
nn.Conv2d(in_channels=64, out_channels=64, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=64),
nn.ReLU(),
nn.MaxPool2d(kernel_size=2, stride=2)
) # [16, 64, 112, 112]
self.conv_block2 = nn.Sequential(
nn.Conv2d(in_channels=64, out_channels=128, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=128),
nn.ReLU(),
nn.Conv2d(in_channels=128, out_channels=128, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=128),
nn.ReLU(),
nn.MaxPool2d(kernel_size=2, stride=2)
) # [16, 128, 56, 56]
self.conv_block3 = nn.Sequential(
nn.Conv2d(in_channels=128, out_channels=256, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=256),
nn.ReLU(),
nn.Conv2d(in_channels=256, out_channels=256, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=256),
nn.ReLU(),
nn.Conv2d(in_channels=256, out_channels=512, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=512),
nn.ReLU(),
nn.MaxPool2d(kernel_size=2, stride=2)
) # [16, 512, 28, 28]
self.conv_block4 = nn.Sequential(
nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=512),
nn.ReLU(),
nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=512),
nn.ReLU(),
nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=512),
nn.ReLU(),
nn.MaxPool2d(kernel_size=2, stride=2)
) # [16, 512, 14, 14]
self.conv_block5 = nn.Sequential(
nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=512),
nn.ReLU(),
nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=512),
nn.ReLU(),
nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(num_features=512),
nn.ReLU(),
nn.MaxPool2d(kernel_size=2, stride=2)
) # [16, 512, 7, 7]
self.classifier = nn.Sequential(
nn.Linear(in_features=512*7*7, out_features=1024),
nn.BatchNorm1d(num_features=1024),
nn.ReLU(),
nn.Linear(in_features=1024, out_features=64),
nn.BatchNorm1d(num_features=64),
nn.ReLU(),
nn.Linear(in_features=64, out_features=2)
)
def forward(self, x):
x = self.conv_block1(x)
x = self.conv_block2(x)
x = self.conv_block3(x)
x = self.conv_block4(x)
x = self.conv_block5(x)
# reshape할 형상 : (batch_size x 512*7*7)
# x = x.view(-1, 512*7*7) # option 1 : view
x = torch.flatten(x, 1) # option 2 : flatten
# x = x.reshape(x.shape[0], -1) # option 3 : reshape
x = self.classifier(x)
return x
model = VGG16()
model.to(device)
model
VGG16(
(conv_block1): Sequential(
(0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU()
(3): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(4): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(5): ReLU()
(6): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(conv_block2): Sequential(
(0): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU()
(3): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(4): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(5): ReLU()
(6): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(conv_block3): Sequential(
(0): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU()
(3): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(4): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(5): ReLU()
(6): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(7): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(8): ReLU()
(9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(conv_block4): Sequential(
(0): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU()
(3): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(4): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(5): ReLU()
(6): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(7): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(8): ReLU()
(9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(conv_block5): Sequential(
(0): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU()
(3): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(4): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(5): ReLU()
(6): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(7): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(8): ReLU()
(9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(linear1): Linear(in_features=25088, out_features=1024, bias=True)
(batch_norm1): BatchNorm1d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(linear2): Linear(in_features=1024, out_features=64, bias=True)
(batch_norm2): BatchNorm1d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(linear3): Linear(in_features=64, out_features=2, bias=True)
)
out = model(images.to(device))
out.shape
torch.Size([16, 2])
for name, parameter in model.named_parameters():
print(name, parameter.size())
5. 모델 설정 (손실함수, 옵티마이저 선택)
# Note. CrossEntropyLoss 관련
# https://pytorch.org/docs/stable/generated/torch.nn.CrossEntropyLoss.html?highlight=crossentropyloss#torch.nn.CrossEntropyLoss
# Note that this case is equivalent to the combination of LogSoftmax and NLLLoss.
# CrossEntropy를 손실함수로 사용하게 되면 forward() 계산시에 softmax() 함수를 사용하면 안됨(otherwise 중복)
# softmax 를 사용하면 부동 소수점 부정확성으로 인해 정확도가 떨어지고 불안정해질 수 있음
# forward()의 마지막 출력은 확률값이 아닌 score(logit)이어야 함
learning_rate = 0.001
# 손실함수
loss_fn = nn.CrossEntropyLoss()
# 옵티마이저(최적화함수, 예:경사하강법)
# optimizer = optim.SGD(model.parameters(), lr=learning_rate)
# 규제의 강도 설정 weight_decay
optimizer = optim.Adam(model.parameters(), lr=learning_rate)
# optimizer = optim.Adam(model.parameters(), lr=learning_rate, weight_decay=0.001)
# Learning Rate Schedule
# https://pytorch.org/docs/stable/generated/torch.optim.lr_scheduler.ReduceLROnPlateau.html
# 모니터링하고 있는 값(예:valid_loss)의 최소값(min) 또는 최대값(max) patience 기간동안 줄어들지 않을 때(OnPlateau) lr에 factor(0.1)를 곱해주는 전략
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', factor=0.1, patience=4, verbose=True)
from torchsummary import summary
# summary(모델, (채널, 인풋사이즈))
summary(model, (3, 224, 224))
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [-1, 64, 224, 224] 1,792
BatchNorm2d-2 [-1, 64, 224, 224] 128
ReLU-3 [-1, 64, 224, 224] 0
Conv2d-4 [-1, 64, 224, 224] 36,928
BatchNorm2d-5 [-1, 64, 224, 224] 128
ReLU-6 [-1, 64, 224, 224] 0
MaxPool2d-7 [-1, 64, 112, 112] 0
Conv2d-8 [-1, 128, 112, 112] 73,856
BatchNorm2d-9 [-1, 128, 112, 112] 256
ReLU-10 [-1, 128, 112, 112] 0
Conv2d-11 [-1, 128, 112, 112] 147,584
BatchNorm2d-12 [-1, 128, 112, 112] 256
ReLU-13 [-1, 128, 112, 112] 0
MaxPool2d-14 [-1, 128, 56, 56] 0
Conv2d-15 [-1, 256, 56, 56] 295,168
BatchNorm2d-16 [-1, 256, 56, 56] 512
ReLU-17 [-1, 256, 56, 56] 0
Conv2d-18 [-1, 256, 56, 56] 590,080
BatchNorm2d-19 [-1, 256, 56, 56] 512
ReLU-20 [-1, 256, 56, 56] 0
Conv2d-21 [-1, 512, 56, 56] 1,180,160
BatchNorm2d-22 [-1, 512, 56, 56] 1,024
ReLU-23 [-1, 512, 56, 56] 0
MaxPool2d-24 [-1, 512, 28, 28] 0
Conv2d-25 [-1, 512, 28, 28] 2,359,808
BatchNorm2d-26 [-1, 512, 28, 28] 1,024
ReLU-27 [-1, 512, 28, 28] 0
Conv2d-28 [-1, 512, 28, 28] 2,359,808
BatchNorm2d-29 [-1, 512, 28, 28] 1,024
ReLU-30 [-1, 512, 28, 28] 0
Conv2d-31 [-1, 512, 28, 28] 2,359,808
BatchNorm2d-32 [-1, 512, 28, 28] 1,024
ReLU-33 [-1, 512, 28, 28] 0
MaxPool2d-34 [-1, 512, 14, 14] 0
Conv2d-35 [-1, 512, 14, 14] 2,359,808
BatchNorm2d-36 [-1, 512, 14, 14] 1,024
ReLU-37 [-1, 512, 14, 14] 0
Conv2d-38 [-1, 512, 14, 14] 2,359,808
BatchNorm2d-39 [-1, 512, 14, 14] 1,024
ReLU-40 [-1, 512, 14, 14] 0
Conv2d-41 [-1, 512, 14, 14] 2,359,808
BatchNorm2d-42 [-1, 512, 14, 14] 1,024
ReLU-43 [-1, 512, 14, 14] 0
MaxPool2d-44 [-1, 512, 7, 7] 0
Linear-45 [-1, 1024] 25,691,136
BatchNorm1d-46 [-1, 1024] 2,048
Linear-47 [-1, 64] 65,600
BatchNorm1d-48 [-1, 64] 128
Linear-49 [-1, 2] 130
================================================================
Total params: 42,252,418
Trainable params: 42,252,418
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.57
Forward/backward pass size (MB): 341.68
Params size (MB): 161.18
Estimated Total Size (MB): 503.43
----------------------------------------------------------------
# 첫번째 conv layer의 모델 파라미터 수
# 필터수 x (필터) + bias
64 * (3*3*3) + 64
1792
# 마지막 출력 feature map의 사이즈
512 * 7 * 7
25088
# linear 1 layer
25088 * 1024 + 1024
25691136
6. 모델 훈련
# torch.no_grad()
# https://pytorch.org/docs/stable/generated/torch.no_grad.html
# Context-manager that disabled gradient calculation.
# Disabling gradient calculation is useful for inference, when you are sure that you will not call Tensor.backward().
# It will reduce memory consumption for computations that would otherwise have requires_grad=True.
def validate(model, validloader, loss_fn):
total = 0
correct = 0
valid_loss = 0
valid_accuracy = 0
# 전방향 예측을 구할 때는 gradient가 필요가 없음음
with torch.no_grad():
for images, labels in validloader: # 이터레이터로부터 next()가 호출되며 미니배치를 반환(images, labels)
# images, labels : (torch.Size([16, 3, 224, 224]), torch.Size([16]))
# 0. Data를 GPU로 보내기
images, labels = images.to(device), labels.to(device)
# 1. 입력 데이터 준비
# not Flatten !!
# images.resize_(images.size()[0], 784)
# 2. 전방향(Forward) 예측
logit = model(images) # 예측 점수
_, preds = torch.max(logit, 1) # 배치에 대한 최종 예측
# preds = logit.max(dim=1)[1]
correct += int((preds == labels).sum()) # 배치 중 맞은 것의 개수가 correct에 누적
total += labels.shape[0] # 배치 사이즈만큼씩 total에 누적
loss = loss_fn(logit, labels)
valid_loss += loss.item() # tensor에서 값을 꺼내와서, 배치의 loss 평균값을 valid_loss에 누적
valid_accuracy = correct / total
return valid_loss, valid_accuracy
# 파이토치에서 텐서보드 사용하기
# https://pytorch.org/tutorials/recipes/recipes/tensorboard_with_pytorch.html
writer = SummaryWriter()
def train_loop(model, trainloader, loss_fn, epochs, optimizer):
steps = 0
steps_per_epoch = len(trainloader)
min_loss = 1000000
max_accuracy = 0
trigger = 0
patience = 7
for epoch in range(epochs):
model.train() # 훈련 모드
train_loss = 0
for images, labels in trainloader: # 이터레이터로부터 next()가 호출되며 미니배치를 반환(images, labels)
steps += 1
# images, labels : (torch.Size([16, 3, 224, 224]), torch.Size([16]))
# 0. Data를 GPU로 보내기
images, labels = images.to(device), labels.to(device)
# 1. 입력 데이터 준비
# not Flatten !!
# images.resize_(images.shape[0], 784)
# 2. 전방향(forward) 예측
predict = model(images) # 예측 점수
loss = loss_fn(predict, labels) # 예측 점수와 정답을 CrossEntropyLoss에 넣어 Loss값 반환
# 3. 역방향(backward) 오차(Gradient) 전파
optimizer.zero_grad() # Gradient가 누적되지 않게 하기 위해
loss.backward() # 모델파리미터들의 Gradient 전파
# 4. 경사 하강법으로 모델 파라미터 업데이트
optimizer.step() # W <- W -lr*Gradient
train_loss += loss.item()
if (steps % steps_per_epoch) == 0 :
model.eval() # 평가 모드 : 평가에서 사용하지 않을 계층(배치 정규화, 드롭아웃)들을 수행하지 않게 하기 위해서
valid_loss, valid_accuracy = validate(model, validloader, loss_fn)
# tensorboard 시각화를 위한 로그 이벤트 등록
writer.add_scalar('Train Loss', train_loss/len(trainloader), epoch+1)
writer.add_scalar('Valid Loss', valid_loss/len(validloader), epoch+1)
writer.add_scalars('Train Loss and Valid Loss',
{'Train' : train_loss/len(trainloader),
'Valid' : valid_loss/len(validloader)}, epoch+1)
writer.add_scalar('Valid Accuracy', valid_accuracy, epoch+1)
# -------------------------------------------
print('Epoch : {}/{}.......'.format(epoch+1, epochs),
'Train Loss : {:.3f}'.format(train_loss/len(trainloader)),
'Valid Loss : {:.3f}'.format(valid_loss/len(validloader)),
'Valid Accuracy : {:.3f}'.format(valid_accuracy)
)
# Best model 저장
# option 1 : valid_loss 모니터링
# if valid_loss < min_loss: # 바로 이전 epoch의 loss보다 작으면 저장하기
# min_loss = valid_loss
# best_model_state = deepcopy(model.state_dict())
# torch.save(best_model_state, 'best_checkpoint.pth')
# option 2 : valid_accuracy 모니터링
if valid_accuracy > max_accuracy : # 바로 이전 epoch의 accuracy보다 크면 저장하기
max_accuracy = valid_accuracy
best_model_state = deepcopy(model.state_dict())
torch.save(best_model_state, 'best_checkpoint.pth')
# -------------------------------------------
# Early Stopping (조기 종료)
if valid_loss > min_loss: # valid_loss가 min_loss를 갱신하지 못하면
trigger += 1
print('trigger : ', trigger)
if trigger > patience:
print('Early Stopping !!!')
print('Training loop is finished !!')
writer.flush()
return
else:
trigger = 0
min_loss = valid_loss
# -------------------------------------------
# Learning Rate Scheduler
scheduler.step(valid_loss)
# -------------------------------------------
writer.flush()
return
epochs = 55
%time train_loop(model, trainloader, loss_fn, epochs, optimizer)
writer.close()
Epoch : 1/55....... Train Loss : 0.698 Valid Loss : 0.685 Valid Accuracy : 0.554
Epoch : 2/55....... Train Loss : 0.685 Valid Loss : 0.660 Valid Accuracy : 0.588
Epoch : 3/55....... Train Loss : 0.666 Valid Loss : 0.647 Valid Accuracy : 0.636
Epoch : 4/55....... Train Loss : 0.660 Valid Loss : 0.643 Valid Accuracy : 0.620
Epoch : 5/55....... Train Loss : 0.633 Valid Loss : 0.703 Valid Accuracy : 0.586
trigger : 1
Epoch : 6/55....... Train Loss : 0.595 Valid Loss : 0.611 Valid Accuracy : 0.660
Epoch : 7/55....... Train Loss : 0.557 Valid Loss : 0.541 Valid Accuracy : 0.732
Epoch : 8/55....... Train Loss : 0.526 Valid Loss : 0.516 Valid Accuracy : 0.733
Epoch : 9/55....... Train Loss : 0.490 Valid Loss : 0.543 Valid Accuracy : 0.725
trigger : 1
Epoch : 10/55....... Train Loss : 0.462 Valid Loss : 0.485 Valid Accuracy : 0.760
Epoch : 11/55....... Train Loss : 0.440 Valid Loss : 0.465 Valid Accuracy : 0.781
Epoch : 12/55....... Train Loss : 0.410 Valid Loss : 0.514 Valid Accuracy : 0.760
trigger : 1
Epoch : 13/55....... Train Loss : 0.386 Valid Loss : 0.467 Valid Accuracy : 0.779
trigger : 2
Epoch : 14/55....... Train Loss : 0.359 Valid Loss : 0.430 Valid Accuracy : 0.799
Epoch : 15/55....... Train Loss : 0.328 Valid Loss : 0.460 Valid Accuracy : 0.802
trigger : 1
Epoch : 16/55....... Train Loss : 0.300 Valid Loss : 0.413 Valid Accuracy : 0.808
Epoch : 17/55....... Train Loss : 0.268 Valid Loss : 0.432 Valid Accuracy : 0.811
trigger : 1
Epoch : 18/55....... Train Loss : 0.250 Valid Loss : 0.515 Valid Accuracy : 0.778
trigger : 2
Epoch : 19/55....... Train Loss : 0.202 Valid Loss : 0.422 Valid Accuracy : 0.824
trigger : 3
Epoch : 20/55....... Train Loss : 0.191 Valid Loss : 0.422 Valid Accuracy : 0.821
trigger : 4
Epoch : 21/55....... Train Loss : 0.165 Valid Loss : 0.458 Valid Accuracy : 0.810
trigger : 5
Epoch 00021: reducing learning rate of group 0 to 1.0000e-04.
Epoch : 22/55....... Train Loss : 0.102 Valid Loss : 0.471 Valid Accuracy : 0.830
trigger : 6
Epoch : 23/55....... Train Loss : 0.067 Valid Loss : 0.431 Valid Accuracy : 0.836
trigger : 7
Epoch : 24/55....... Train Loss : 0.061 Valid Loss : 0.429 Valid Accuracy : 0.836
trigger : 8
Early Stopping !!!
Training loop is finished !!
CPU times: user 1h 18min 13s, sys: 20 s, total: 1h 18min 33s
Wall time: 1h 2min 54s
%load_ext tensorboard
%tensorboard --logdir=runs
Output hidden; open in https://colab.research.google.com to view.
7. 모델 예측
# testloader에서 미니 배치 가져오기
test_iter = iter(testloader)
images, labels = next(test_iter)
images, labels = images.to(device), labels.to(device)
print(images.size(), labels.size())
# random한 index로 이미지 한장 준비하기
rnd_idx = 10
print(images[rnd_idx].shape, labels[rnd_idx])
torch.Size([16, 3, 224, 224]) torch.Size([16])
torch.Size([3, 224, 224]) tensor(0, device='cuda:0')
images[rnd_idx].shape
torch.Size([3, 224, 224])
# not Flatten!
# flattend_img = images[rnd_idx].view(1, 784)
# 준비된 이미지로 예측하기
model.eval()
with torch.no_grad():
logit = model(images[rnd_idx].unsqueeze(0)) # model.forward()에서는 배치가 적용된 4차원 입력 기대
pred = logit.max(dim=1)[1]
print(pred == labels[rnd_idx]) # True : 잘 예측
tensor([True], device='cuda:0')
print("pred:", pred, "labels:", labels[rnd_idx])
print(labels_map[pred.cpu().item()], labels_map[labels[rnd_idx].cpu().item()])
plt.imshow(images[rnd_idx].permute(1, 2, 0).cpu())
pred: tensor([0], device='cuda:0') labels: tensor(0, device='cuda:0')
cat cat
<matplotlib.image.AxesImage at 0x7f47e6df7730>
 VGGNet)/output_73_2.png)
8. 모델 평가
def evaluation(model, testloader, loss_fn):
total = 0
correct = 0
test_loss = 0
test_accuracy = 0
# 전방향 예측을 구할 때는 gradient가 필요가 없음음
with torch.no_grad():
for images, labels in testloader: # 이터레이터로부터 next()가 호출되며 미니배치를 반환(images, labels)
# 0. Data를 GPU로 보내기
images, labels = images.to(device), labels.to(device)
# 1. 입력 데이터 준비
# not Flatten
# images.resize_(images.size()[0], 784)
# 2. 전방향(Forward) 예측
logit = model(images) # 예측 점수
_, preds = torch.max(logit, 1) # 배치에 대한 최종 예측
# preds = logit.max(dim=1)[1]
correct += int((preds == labels).sum()) # 배치치 중 맞은 것의 개수가 correct에 누적
total += labels.shape[0] # 배치 사이즈만큼씩 total에 누적
loss = loss_fn(logit, labels)
test_loss += loss.item() # tensor에서 값을 꺼내와서, 배치의 loss 평균값을 valid_loss에 누적
test_accuracy = correct / total
print('Test Loss : {:.3f}'.format(test_loss/len(testloader)),
'Test Accuracy : {:.3f}'.format(test_accuracy))
model.eval()
evaluation(model, testloader, loss_fn)
Test Loss : 0.434 Test Accuracy : 0.834
9. 모델 저장
# 모델을 저장하는 이유
# 1. 예측을 할 때마다 훈련시키는것은 비효율적
# 2. 기존 훈련 결과에 이어서 학습을 하고자 할 때
# 파이토치에서 모델 저장하기
# https://pytorch.org/tutorials/beginner/saving_loading_models.html
# 현재 모델에 저장되어 있는 모델 파라미터터
model.state_dict().keys()
odict_keys(['conv_block1.0.weight', 'conv_block1.0.bias', 'conv_block1.1.weight', 'conv_block1.1.bias', 'conv_block1.1.running_mean', 'conv_block1.1.running_var', 'conv_block1.1.num_batches_tracked', 'conv_block1.3.weight', 'conv_block1.3.bias', 'conv_block1.4.weight', 'conv_block1.4.bias', 'conv_block1.4.running_mean', 'conv_block1.4.running_var', 'conv_block1.4.num_batches_tracked', 'conv_block2.0.weight', 'conv_block2.0.bias', 'conv_block2.1.weight', 'conv_block2.1.bias', 'conv_block2.1.running_mean', 'conv_block2.1.running_var', 'conv_block2.1.num_batches_tracked', 'conv_block2.3.weight', 'conv_block2.3.bias', 'conv_block2.4.weight', 'conv_block2.4.bias', 'conv_block2.4.running_mean', 'conv_block2.4.running_var', 'conv_block2.4.num_batches_tracked', 'conv_block3.0.weight', 'conv_block3.0.bias', 'conv_block3.1.weight', 'conv_block3.1.bias', 'conv_block3.1.running_mean', 'conv_block3.1.running_var', 'conv_block3.1.num_batches_tracked', 'conv_block3.3.weight', 'conv_block3.3.bias', 'conv_block3.4.weight', 'conv_block3.4.bias', 'conv_block3.4.running_mean', 'conv_block3.4.running_var', 'conv_block3.4.num_batches_tracked', 'conv_block3.6.weight', 'conv_block3.6.bias', 'conv_block3.7.weight', 'conv_block3.7.bias', 'conv_block3.7.running_mean', 'conv_block3.7.running_var', 'conv_block3.7.num_batches_tracked', 'conv_block4.0.weight', 'conv_block4.0.bias', 'conv_block4.1.weight', 'conv_block4.1.bias', 'conv_block4.1.running_mean', 'conv_block4.1.running_var', 'conv_block4.1.num_batches_tracked', 'conv_block4.3.weight', 'conv_block4.3.bias', 'conv_block4.4.weight', 'conv_block4.4.bias', 'conv_block4.4.running_mean', 'conv_block4.4.running_var', 'conv_block4.4.num_batches_tracked', 'conv_block4.6.weight', 'conv_block4.6.bias', 'conv_block4.7.weight', 'conv_block4.7.bias', 'conv_block4.7.running_mean', 'conv_block4.7.running_var', 'conv_block4.7.num_batches_tracked', 'conv_block5.0.weight', 'conv_block5.0.bias', 'conv_block5.1.weight', 'conv_block5.1.bias', 'conv_block5.1.running_mean', 'conv_block5.1.running_var', 'conv_block5.1.num_batches_tracked', 'conv_block5.3.weight', 'conv_block5.3.bias', 'conv_block5.4.weight', 'conv_block5.4.bias', 'conv_block5.4.running_mean', 'conv_block5.4.running_var', 'conv_block5.4.num_batches_tracked', 'conv_block5.6.weight', 'conv_block5.6.bias', 'conv_block5.7.weight', 'conv_block5.7.bias', 'conv_block5.7.running_mean', 'conv_block5.7.running_var', 'conv_block5.7.num_batches_tracked', 'linear1.weight', 'linear1.bias', 'batch_norm1.weight', 'batch_norm1.bias', 'batch_norm1.running_mean', 'batch_norm1.running_var', 'batch_norm1.num_batches_tracked', 'linear2.weight', 'linear2.bias', 'batch_norm2.weight', 'batch_norm2.bias', 'batch_norm2.running_mean', 'batch_norm2.running_var', 'batch_norm2.num_batches_tracked', 'linear3.weight', 'linear3.bias'])
torch.save(model.state_dict(), 'last_checkpoint.pth')
# 시간이 흐른뒤 다시 모델 가져오기
last_state_dict = torch.load('last_checkpoint.pth')
last_state_dict.keys()
odict_keys(['conv_block1.0.weight', 'conv_block1.0.bias', 'conv_block1.1.weight', 'conv_block1.1.bias', 'conv_block1.1.running_mean', 'conv_block1.1.running_var', 'conv_block1.1.num_batches_tracked', 'conv_block1.3.weight', 'conv_block1.3.bias', 'conv_block1.4.weight', 'conv_block1.4.bias', 'conv_block1.4.running_mean', 'conv_block1.4.running_var', 'conv_block1.4.num_batches_tracked', 'conv_block2.0.weight', 'conv_block2.0.bias', 'conv_block2.1.weight', 'conv_block2.1.bias', 'conv_block2.1.running_mean', 'conv_block2.1.running_var', 'conv_block2.1.num_batches_tracked', 'conv_block2.3.weight', 'conv_block2.3.bias', 'conv_block2.4.weight', 'conv_block2.4.bias', 'conv_block2.4.running_mean', 'conv_block2.4.running_var', 'conv_block2.4.num_batches_tracked', 'conv_block3.0.weight', 'conv_block3.0.bias', 'conv_block3.1.weight', 'conv_block3.1.bias', 'conv_block3.1.running_mean', 'conv_block3.1.running_var', 'conv_block3.1.num_batches_tracked', 'conv_block3.3.weight', 'conv_block3.3.bias', 'conv_block3.4.weight', 'conv_block3.4.bias', 'conv_block3.4.running_mean', 'conv_block3.4.running_var', 'conv_block3.4.num_batches_tracked', 'conv_block3.6.weight', 'conv_block3.6.bias', 'conv_block3.7.weight', 'conv_block3.7.bias', 'conv_block3.7.running_mean', 'conv_block3.7.running_var', 'conv_block3.7.num_batches_tracked', 'conv_block4.0.weight', 'conv_block4.0.bias', 'conv_block4.1.weight', 'conv_block4.1.bias', 'conv_block4.1.running_mean', 'conv_block4.1.running_var', 'conv_block4.1.num_batches_tracked', 'conv_block4.3.weight', 'conv_block4.3.bias', 'conv_block4.4.weight', 'conv_block4.4.bias', 'conv_block4.4.running_mean', 'conv_block4.4.running_var', 'conv_block4.4.num_batches_tracked', 'conv_block4.6.weight', 'conv_block4.6.bias', 'conv_block4.7.weight', 'conv_block4.7.bias', 'conv_block4.7.running_mean', 'conv_block4.7.running_var', 'conv_block4.7.num_batches_tracked', 'conv_block5.0.weight', 'conv_block5.0.bias', 'conv_block5.1.weight', 'conv_block5.1.bias', 'conv_block5.1.running_mean', 'conv_block5.1.running_var', 'conv_block5.1.num_batches_tracked', 'conv_block5.3.weight', 'conv_block5.3.bias', 'conv_block5.4.weight', 'conv_block5.4.bias', 'conv_block5.4.running_mean', 'conv_block5.4.running_var', 'conv_block5.4.num_batches_tracked', 'conv_block5.6.weight', 'conv_block5.6.bias', 'conv_block5.7.weight', 'conv_block5.7.bias', 'conv_block5.7.running_mean', 'conv_block5.7.running_var', 'conv_block5.7.num_batches_tracked', 'linear1.weight', 'linear1.bias', 'batch_norm1.weight', 'batch_norm1.bias', 'batch_norm1.running_mean', 'batch_norm1.running_var', 'batch_norm1.num_batches_tracked', 'linear2.weight', 'linear2.bias', 'batch_norm2.weight', 'batch_norm2.bias', 'batch_norm2.running_mean', 'batch_norm2.running_var', 'batch_norm2.num_batches_tracked', 'linear3.weight', 'linear3.bias'])
# 읽어들인 모델 파라미터는 모델 아키텍처에 연결을 시켜줘야 함
# load_state_dict() 사용
last_model = VGG16()
last_model.to(device)
last_model.load_state_dict(last_state_dict)
<All keys matched successfully>
last_model.eval()
evaluation(last_model, testloader, loss_fn)
Test Loss : 0.434 Test Accuracy : 0.834
# valid loss or accuracy 기준 best model
best_state_dict = torch.load('best_checkpoint.pth')
best_model = VGG16()
best_model.to(device)
best_model.load_state_dict(best_state_dict)
<All keys matched successfully>
best_model.eval()
evaluation(best_model, testloader, loss_fn)
Test Loss : 0.425 Test Accuracy : 0.840
#best_state_dict['conv_block1.0.weight']
#last_state_dict['conv_block1.0.weight']
댓글남기기