- 🍨 本文为🔗365天深度学习训练营 中的学习记录博客
- 🍖 原作者：K同学啊
前言
- 本来想继续优化的，但是我看论文和查阅一些资料，涉及到了知识蒸馏、量化的知识，这些知识我需要花一点时间去研究一下，才能进一步优化。
🍺 实验要求：
- 自己搭建VGG-16网络框架
- 如何查看模型的参数量以及相关指标
🍻 拔高（可选）：
- 验证集准确率达到100%
- 使用PPT画出VGG-16算法框架图
🔎 探索（难度有点大）
- 在不影响准确率的前提下轻量化模型
- 目前VGG16的Total params是134,272,835
🏩 完成度：
- 三个任务均已完成
- 自己搭建试验成果：训练集准确率100%，测试集准确率99.1%
- 提高：分别对分类层、卷积层进行了优化，但是结果却没有啥变化
- 探索：全面修改分类层，计算量减少了99%，但是
Train_acc:95.8%,Test_acc:94.9%，准确率降低了4%左右
📟 论文参考
[1]方宇伦,陈雪纯,杜世昌,等.基于轻量化深度学习VGG16网络模型的表面缺陷检测方法[J].机械设计与研究,2023,39(02):143-147.DOI:10.13952/j.cnki.jofmdr.2023.0068.
文章目录
- vgg16算法简介
- 实验验证
- 1、数据处理
- 1、导入库
- 2、查看文件要分类的类型
- 3、展示图片
- 4、导入全部数据与数据处理
- 5、数据划分
- 6、动态加载数据
- 2、VGG16模型的搭建
- 3、模型训练
- 1、模型训练搭建
- 2、模型测试构建
- 3、设置超参数
- 4、正式训练
- 4、结果可视化
- 5、优化
- 优化一
- 优化二
- 6、轻量级优化
- 优化前的分类层
- 参数计算
- 总参数数量
- 优化后的分类层
- 参数计算
- 总参数数量
- 轻量化程度
- 参数减少比例
- 计算复杂度
vgg16算法简介
vgg16是CNN卷积神经网络的经典架构，它拥有13层卷积，3层池化构成，下图是本人第一次画的神经网络🤠🤠🤠🤠
实验验证
1、数据处理
1、导入库
import torch
import torchvision
import torch.nn as nn
import numpy as np
device = ('cuda' if torch.cuda.is_available() else 'cpu')
device
输出：
'cuda'
2、查看文件要分类的类型
import os, PIL, pathlib
data_dir = './data/'
data_dir = pathlib.Path(data_dir) # 转换为 pathlib 类型
# 查看./data/下的文件夹名称
data_path = data_dir.glob('*')
classNames = [str(path).split('/')[1] for path in data_path]
classNames
输出：
['Early_blight', 'healthy', 'Late_blight']
3、展示图片
import matplotlib.pyplot as plt
from PIL import Image
# 获取图片路径名，展示早衰图片
data_look_dir = './data/Early_blight/'
data_path_list = [f for f in os.listdir(data_look_dir) if f.endswith(('JPG', 'png'))]
# 创建画板
fig, axes = plt.subplots(2, 8, figsize=(16, 4))
# 图片展示
for ax, img_name in zip(axes.flat, data_path_list):
path_name = os.path.join(data_look_dir, img_name)
img = Image.open(path_name)
ax.imshow(img)
ax.axis('off')
plt.show()
4、导入全部数据与数据处理
from torchvision import transforms, datasets
# 图片处理--> 统一
data_transforms = transforms.Compose([
transforms.Resize([224, 224]),
transforms.ToTensor(),
transforms.Normalize( # 归一化
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]
)
])
# datasets.ImageFolder(): 专门加载图像分类的API
total_dir = './data/'
total_data = datasets.ImageFolder(root=total_dir, transform=data_transforms)
5、数据划分
train_size = int(len(total_data) * 0.8)
test_size = len(total_data) - train_size
# 随机划分
train_dataset, test_dataset = torch.utils.data.random_split(total_data, [train_size, test_size])
print(train_dataset)
print(test_dataset)
<torch.utils.data.dataset.Subset object at 0x7f17d1326e80>
<torch.utils.data.dataset.Subset object at 0x7f18e029db20>
6、动态加载数据
batch_size = 32
train_dl = torch.utils.data.DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=1)
test_dl = torch.utils.data.DataLoader(test_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=1)
# 展示一个加载后的数据格式
for X, data in test_dl:
print("shape: ", X.shape)
print("data: ", data)
break
shape: torch.Size([32, 3, 224, 224])
data: tensor([1, 1, 0, 0, 0, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 0, 1, 2, 0, 1, 1, 1, 1, 1,
0, 0, 1, 1, 0, 1, 2, 0])
2、VGG16模型的搭建
class vgg16_model(nn.Module):
def __init__(self):
super(vgg16_model, self).__init__()
# block1
self.block1 = nn.Sequential(
nn.Conv2d(in_channels=3, out_channels=64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
nn.BatchNorm2d(64),
nn.ReLU(),
nn.Conv2d(in_channels=64, out_channels=64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
nn.BatchNorm2d(64),
nn.ReLU(),
nn.MaxPool2d(kernel_size=(2, 2), stride=(2, 2))
)
# block2
self.block2 = nn.Sequential(
nn.Conv2d(in_channels=64, out_channels=128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
nn.ReLU(),
nn.Conv2d(in_channels=128, out_channels=128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
nn.ReLU(),
nn.MaxPool2d(kernel_size=(2, 2), stride=(2, 2))
)
# block3
self.block3 = nn.Sequential(
nn.Conv2d(in_channels=128, out_channels=256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
nn.ReLU(),
nn.Conv2d(in_channels=256, out_channels=256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
nn.ReLU(),
nn.Conv2d(in_channels=256, out_channels=256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
nn.ReLU(),
nn.MaxPool2d(kernel_size=(2, 2), stride=(2, 2))
)
# block4
self.block4 = nn.Sequential(
nn.Conv2d(in_channels=256, out_channels=512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
nn.ReLU(),
nn.Conv2d(in_channels=512, out_channels=512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
nn.ReLU(),
nn.Conv2d(in_channels=512, out_channels=512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
nn.ReLU(),
nn.MaxPool2d(kernel_size=(2, 2), stride=(2, 2))
)
# block5
self.block5 = nn.Sequential(
nn.Conv2d(in_channels=512, out_channels=512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
nn.ReLU(),
nn.Conv2d(in_channels=512, out_channels=512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
nn.ReLU(),
nn.Conv2d(in_channels=512, out_channels=512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
nn.ReLU(),
nn.MaxPool2d(kernel_size=(2, 2), stride=(2, 2))
)
# classifier
self.classifier = nn.Sequential(
nn.Linear(in_features=512 * 7 * 7, out_features=4096),
nn.ReLU(),
nn.Linear(in_features=4096, out_features=4096),
nn.ReLU(),
nn.Linear(in_features=4096, out_features=3)
)
def forward(self, x):
x = self.block1(x)
x = self.block2(x)
x = self.block3(x)
x = self.block4(x)
x = self.block5(x)
x = torch.flatten(x, start_dim=1)
x = self.classifier(x)
return x
model = vgg16_model().to(device)
model
输出：
vgg16_model(
(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, 2), stride=(2, 2), padding=0, dilation=1, ceil_mode=False)
)
(block2): Sequential(
(0): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU()
(2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU()
(4): MaxPool2d(kernel_size=(2, 2), stride=(2, 2), padding=0, dilation=1, ceil_mode=False)
)
(block3): Sequential(
(0): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU()
(2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU()
(4): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(5): ReLU()
(6): MaxPool2d(kernel_size=(2, 2), stride=(2, 2), padding=0, dilation=1, ceil_mode=False)
)
(block4): Sequential(
(0): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU()
(2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU()
(4): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(5): ReLU()
(6): MaxPool2d(kernel_size=(2, 2), stride=(2, 2), padding=0, dilation=1, ceil_mode=False)
)
(block5): Sequential(
(0): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU()
(2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU()
(4): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(5): ReLU()
(6): MaxPool2d(kernel_size=(2, 2), stride=(2, 2), padding=0, dilation=1, ceil_mode=False)
)
(classifier): Sequential(
(0): Linear(in_features=25088, out_features=4096, bias=True)
(1): ReLU()
(2): Linear(in_features=4096, out_features=4096, bias=True)
(3): ReLU()
(4): Linear(in_features=4096, out_features=3, bias=True)
)
)
继续输出模型详细参数：
# 统计模型的参数以及其他指标
import torchsummary as summary
summary.summary(model, (3, 244, 244))
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [-1, 64, 244, 244] 1,792
BatchNorm2d-2 [-1, 64, 244, 244] 128
ReLU-3 [-1, 64, 244, 244] 0
Conv2d-4 [-1, 64, 244, 244] 36,928
BatchNorm2d-5 [-1, 64, 244, 244] 128
ReLU-6 [-1, 64, 244, 244] 0
MaxPool2d-7 [-1, 64, 122, 122] 0
Conv2d-8 [-1, 128, 122, 122] 73,856
ReLU-9 [-1, 128, 122, 122] 0
Conv2d-10 [-1, 128, 122, 122] 147,584
ReLU-11 [-1, 128, 122, 122] 0
MaxPool2d-12 [-1, 128, 61, 61] 0
Conv2d-13 [-1, 256, 61, 61] 295,168
ReLU-14 [-1, 256, 61, 61] 0
Conv2d-15 [-1, 256, 61, 61] 590,080
ReLU-16 [-1, 256, 61, 61] 0
Conv2d-17 [-1, 256, 61, 61] 590,080
ReLU-18 [-1, 256, 61, 61] 0
MaxPool2d-19 [-1, 256, 30, 30] 0
Conv2d-20 [-1, 512, 30, 30] 1,180,160
ReLU-21 [-1, 512, 30, 30] 0
Conv2d-22 [-1, 512, 30, 30] 2,359,808
ReLU-23 [-1, 512, 30, 30] 0
Conv2d-24 [-1, 512, 30, 30] 2,359,808
ReLU-25 [-1, 512, 30, 30] 0
MaxPool2d-26 [-1, 512, 15, 15] 0
Conv2d-27 [-1, 512, 15, 15] 2,359,808
ReLU-28 [-1, 512, 15, 15] 0
Conv2d-29 [-1, 512, 15, 15] 2,359,808
ReLU-30 [-1, 512, 15, 15] 0
Conv2d-31 [-1, 512, 15, 15] 2,359,808
ReLU-32 [-1, 512, 15, 15] 0
MaxPool2d-33 [-1, 512, 7, 7] 0
Linear-34 [-1, 4096] 102,764,544
ReLU-35 [-1, 4096] 0
Linear-36 [-1, 4096] 16,781,312
ReLU-37 [-1, 4096] 0
Linear-38 [-1, 3] 12,291
================================================================
Total params: 134,273,091
Trainable params: 134,273,091
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.68
Forward/backward pass size (MB): 316.39
Params size (MB): 512.21
Estimated Total Size (MB): 829.28
----------------------------------------------------------------
3、模型训练
1、模型训练搭建
def train(dataloader, model, loss_fn, optimizer):
size = len(dataloader.dataset)
batch_size = len(dataloader)
train_acc, train_loss = 0, 0
for X, y in dataloader:
X, y = X.to(device), y.to(device)
# 训练
pred = model(X)
loss = loss_fn(pred, y)
# 梯度下降法
optimizer.zero_grad()
loss.backward()
optimizer.step()
# 记录
train_loss += loss.item()
train_acc += (pred.argmax(1) == y).type(torch.float).sum().item()
train_acc /= size
train_loss /= batch_size
return train_acc, train_loss
2、模型测试构建
def test(dataloader, model, loss_fn):
size = len(dataloader.dataset)
batch_size = len(dataloader)
test_acc, test_loss = 0, 0
with torch.no_grad():
for X, y in dataloader:
X, y = X.to(device), y.to(device)
pred = model(X)
loss = loss_fn(pred, y)
test_loss += loss.item()
test_acc += (pred.argmax(1) == y).type(torch.float).sum().item()
test_acc /= size
test_loss /= batch_size
return test_acc, test_loss
3、设置超参数
loss_fn = nn.CrossEntropyLoss() # 损失函数
learn_lr = 1e-4 # 超参数
optimizer = torch.optim.Adam(model.parameters(), lr=learn_lr) # 优化器
4、正式训练
train_acc = []
train_loss = []
test_acc = []
test_loss = []
epoches = 40
for i in range(epoches):
model.train()
epoch_train_acc, epoch_train_loss = train(train_dl, model, loss_fn, optimizer)
model.eval()
epoch_test_acc, epoch_test_loss = test(test_dl, model, loss_fn)
train_acc.append(epoch_train_acc)
train_loss.append(epoch_train_loss)
test_acc.append(epoch_test_acc)
test_loss.append(epoch_test_loss)
# 输出
template = ('Epoch:{:2d}, Train_acc:{:.1f}%, Train_loss:{:.3f}, Test_acc:{:.1f}%, Test_loss:{:.3f}')
print(template.format(i + 1, epoch_train_acc*100, epoch_train_loss, epoch_test_acc*100, epoch_test_loss))
print("Done")
Epoch: 1, Train_acc:45.4%, Train_loss:0.927, Test_acc:51.7%, Test_loss:0.909
Epoch: 2, Train_acc:63.6%, Train_loss:0.744, Test_acc:85.4%, Test_loss:0.491
Epoch: 3, Train_acc:83.4%, Train_loss:0.456, Test_acc:82.8%, Test_loss:0.423
Epoch: 4, Train_acc:87.3%, Train_loss:0.340, Test_acc:90.3%, Test_loss:0.276
Epoch: 5, Train_acc:89.8%, Train_loss:0.265, Test_acc:87.9%, Test_loss:0.294
Epoch: 6, Train_acc:89.8%, Train_loss:0.260, Test_acc:95.6%, Test_loss:0.136
Epoch: 7, Train_acc:95.4%, Train_loss:0.122, Test_acc:97.4%, Test_loss:0.087
Epoch: 8, Train_acc:95.9%, Train_loss:0.120, Test_acc:94.0%, Test_loss:0.164
Epoch: 9, Train_acc:97.4%, Train_loss:0.080, Test_acc:94.7%, Test_loss:0.117
Epoch:10, Train_acc:96.2%, Train_loss:0.102, Test_acc:97.2%, Test_loss:0.082
Epoch:11, Train_acc:95.0%, Train_loss:0.134, Test_acc:97.9%, Test_loss:0.083
Epoch:12, Train_acc:96.6%, Train_loss:0.093, Test_acc:96.8%, Test_loss:0.069
Epoch:13, Train_acc:98.2%, Train_loss:0.057, Test_acc:97.9%, Test_loss:0.063
Epoch:14, Train_acc:98.4%, Train_loss:0.059, Test_acc:98.4%, Test_loss:0.047
Epoch:15, Train_acc:98.4%, Train_loss:0.043, Test_acc:97.4%, Test_loss:0.058
Epoch:16, Train_acc:98.0%, Train_loss:0.059, Test_acc:95.4%, Test_loss:0.107
Epoch:17, Train_acc:97.5%, Train_loss:0.065, Test_acc:96.3%, Test_loss:0.088
Epoch:18, Train_acc:98.8%, Train_loss:0.037, Test_acc:98.1%, Test_loss:0.049
Epoch:19, Train_acc:98.6%, Train_loss:0.037, Test_acc:98.1%, Test_loss:0.049
Epoch:20, Train_acc:98.4%, Train_loss:0.043, Test_acc:97.9%, Test_loss:0.065
Epoch:21, Train_acc:99.5%, Train_loss:0.010, Test_acc:94.2%, Test_loss:0.300
Epoch:22, Train_acc:97.9%, Train_loss:0.069, Test_acc:93.7%, Test_loss:0.220
Epoch:23, Train_acc:99.4%, Train_loss:0.025, Test_acc:97.9%, Test_loss:0.085
Epoch:24, Train_acc:99.4%, Train_loss:0.019, Test_acc:97.0%, Test_loss:0.090
Epoch:25, Train_acc:98.8%, Train_loss:0.033, Test_acc:98.4%, Test_loss:0.060
Epoch:26, Train_acc:99.4%, Train_loss:0.018, Test_acc:97.0%, Test_loss:0.125
Epoch:27, Train_acc:98.9%, Train_loss:0.031, Test_acc:97.7%, Test_loss:0.091
Epoch:28, Train_acc:99.8%, Train_loss:0.005, Test_acc:97.4%, Test_loss:0.080
Epoch:29, Train_acc:99.8%, Train_loss:0.006, Test_acc:94.4%, Test_loss:0.360
Epoch:30, Train_acc:97.7%, Train_loss:0.058, Test_acc:97.9%, Test_loss:0.050
Epoch:31, Train_acc:98.6%, Train_loss:0.034, Test_acc:98.1%, Test_loss:0.065
Epoch:32, Train_acc:99.7%, Train_loss:0.014, Test_acc:98.1%, Test_loss:0.056
Epoch:33, Train_acc:99.4%, Train_loss:0.013, Test_acc:99.3%, Test_loss:0.047
Epoch:34, Train_acc:99.5%, Train_loss:0.020, Test_acc:98.6%, Test_loss:0.069
Epoch:35, Train_acc:99.4%, Train_loss:0.021, Test_acc:98.4%, Test_loss:0.055
Epoch:36, Train_acc:99.7%, Train_loss:0.010, Test_acc:98.1%, Test_loss:0.080
Epoch:37, Train_acc:99.7%, Train_loss:0.010, Test_acc:98.4%, Test_loss:0.059
Epoch:38, Train_acc:99.3%, Train_loss:0.017, Test_acc:98.4%, Test_loss:0.055
Epoch:39, Train_acc:99.3%, Train_loss:0.017, Test_acc:98.8%, Test_loss:0.043
Epoch:40, Train_acc:99.9%, Train_loss:0.004, Test_acc:98.8%, Test_loss:0.038
Done
我第一次、第二次跑的训练集准确率达到了100%，这是后面跑的，也正常，因为深度学习每一次更新的梯度、权重、偏置等都不一样。
4、结果可视化
import matplotlib.pyplot as plt
#隐藏警告
import warnings
warnings.filterwarnings("ignore") #忽略警告信息
epochs_range = range(epoches)
plt.figure(figsize=(12, 3))
plt.subplot(1, 2, 1)
plt.plot(epochs_range, train_acc, label='Training Accuracy')
plt.plot(epochs_range, test_acc, label='Test Accuracy')
plt.legend(loc='lower right')
plt.title('Training Accuracy')
plt.subplot(1, 2, 2)
plt.plot(epochs_range, train_loss, label='Training Loss')
plt.plot(epochs_range, test_loss, label='Test Loss')
plt.legend(loc='upper right')
plt.title('Training= Loss')
plt.show()
- 但从效果来看，这个以及够了，极好了，没什么必要提高了，但是从我初学的角度上来看，提高优化以下还是很有必要的
5、优化
两次优化都没啥提升，大差不差，有些反而提高了计算量
优化一
在每一层卷积层后添加BN层，提高稳定性，结果如下：
优化二
在分类层中，降低分类层数量，512*7*7 —> 1024 —> 512 --> 3，添加Dropout层，这个思路效果大差不大
6、轻量级优化
优化：全面修改全连接层，首先添加平均池化层，后添加Dropout层、Linear(512, 3)，本来还想给添加L1正则化的，但是效果极差。
结果：Train_acc:95.8%, Test_acc:94.9%，但是，参数数量减少了约 99.9955%，效果极好
分析过程：
通过优化 VGG16 的分类层，您显著减少了模型的参数数量和计算复杂度。下面是对优化前后模型的对比分析：
优化前的分类层
self.classifier = nn.Sequential(
nn.Linear(in_features=512 * 7 * 7, out_features=4096),
nn.ReLU(),
nn.Linear(in_features=4096, out_features=4096),
nn.ReLU(),
nn.Linear(in_features=4096, out_features=3)
)
参数计算
- 第一层：
nn.Linear(in_features=512 * 7 * 7, out_features=4096)- 输入大小：512 * 7 * 7 = 25088
- 输出大小：4096
- 参数数量：25088 * 4096 + 4096 = 103219200 + 4096 = 103223296
- 第二层：
nn.Linear(in_features=4096, out_features=4096)- 输入大小：4096
- 输出大小：4096
- 参数数量：4096 * 4096 + 4096 = 16781312 + 4096 = 16785408
- 第三层：
nn.Linear(in_features=4096, out_features=3)- 输入大小：4096
- 输出大小：3
- 参数数量：4096 * 3 + 3 = 12288 + 3 = 12291
总参数数量
- 总参数数量：103223296 + 16785408 + 12291 = 120020995
优化后的分类层
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.classifier = nn.Sequential(
nn.Dropout(),
nn.Linear(512, 3)
)
参数计算
- 全局平均池化层：
nn.AdaptiveAvgPool2d((1, 1))- 不增加参数，只是将特征图的大小从 7x7 压缩到 1x1。
- 线性层：
nn.Linear(512, 3)- 输入大小：512
- 输出大小：3
- 参数数量：512 * 3 + 3 = 1536 + 3 = 1539
总参数数量
- 总参数数量：1539
轻量化程度
- 优化前的参数数量：120020995
- 优化后的参数数量：1539
参数减少比例
- 参数减少比例：(120020995 - 1539) / 120020995 ≈ 0.999955
- 即优化后的模型参数数量减少了约 99.9955%。
计算复杂度
- 优化前：涉及多个全连接层，计算量较大。
- 优化后：只涉及一个全连接层，计算量显著减少。
