第P9周：YOLOv5-Backbone模块实现
- 🍨 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
环境：
- 编程语言：Python(3.10.12)
- 编译器：Google Colab
- 数据集：天气识别数据
- 深度学习环境：pytorch
- torch == 2.3.0+cu121
- torchvision == 0.18.0+cu121
文章目录
一、前期准备
1.设置环境
import torch
import torch.nn as nn
import torchvision.transforms as transforms
import torchvision
from torchvision import transforms, datasets
import os,PIL,pathlib,warnings
warnings.filterwarnings("ignore") #忽略警告信息
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
device
device(type='cuda')
2.导入数据
from google.colab import drive
drive.mount("/content/drive/")
Mounted at /content/drive/
%cd "/content/drive/MyDrive/Colab Notebooks/jupyter notebook/data/"
/content/drive/Othercomputers/My laptop/jupyter notebook/data
import os,PIL,random,pathlib
data_dir = './weather_photos/'
data_dir = pathlib.Path(data_dir)
data_paths = list(data_dir.glob('*'))
classNames = [str(path).split("/")[1] for path in data_paths]
classNames
['rain', 'shine', 'sunrise', 'cloudy']
# 关于transforms.Compose的更多介绍可以参考：https://blog.csdn.net/qq_38251616/article/details/124878863
train_transforms = transforms.Compose([
transforms.Resize([224, 224]), # 将输入图片resize成统一尺寸
# transforms.RandomHorizontalFlip(), # 随机水平翻转
transforms.ToTensor(), # 将PIL Image或numpy.ndarray转换为tensor，并归一化到[0,1]之间
transforms.Normalize( # 标准化处理-->转换为标准正太分布（高斯分布），使模型更容易收敛
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]) # 其中 mean=[0.485,0.456,0.406]与std=[0.229,0.224,0.225] 从数据集中随机抽样计算得到的。
])
test_transform = transforms.Compose([
transforms.Resize([224, 224]), # 将输入图片resize成统一尺寸
transforms.ToTensor(), # 将PIL Image或numpy.ndarray转换为tensor，并归一化到[0,1]之间
transforms.Normalize( # 标准化处理-->转换为标准正太分布（高斯分布），使模型更容易收敛
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]) # 其中 mean=[0.485,0.456,0.406]与std=[0.229,0.224,0.225] 从数据集中随机抽样计算得到的。
])
total_data = datasets.ImageFolder("./weather_photos/",transform=train_transforms)
total_data
Dataset ImageFolder
Number of datapoints: 1125
Root location: ./weather_photos/
StandardTransform
Transform: Compose(
Resize(size=[224, 224], interpolation=bilinear, max_size=None, antialias=True)
ToTensor()
Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
)
total_data.class_to_idx
{'cloudy': 0, 'rain': 1, 'shine': 2, 'sunrise': 3}
3.划分数据集
train_size = int(0.8*len(total_data))
test_size = len(total_data) - train_size
train_dataset, test_dataset = torch.utils.data.random_split(total_data,[train_size, test_size])
train_dataset, test_dataset
(<torch.utils.data.dataset.Subset at 0x7987b42a48e0>,
<torch.utils.data.dataset.Subset at 0x7987b42a56c0>)
batch_size = 8
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)
for X, y in test_dl:
print("Shape of X [N, C, H, W]: ", X.shape)
print("Shape of y: ", y.shape, y.dtype)
break
Shape of X [N, C, H, W]: torch.Size([8, 3, 224, 224])
Shape of y: torch.Size([8]) torch.int64
二、搭建含Backbone模块的模型
1.YOLOv5简介:
以下博客均较为详细介绍了YOLOv5的结构：
- yolov5原理详解 (涉及内容：Yolov5框架，各组件分析，特征融合是怎么实现的？yolov5的具体特征融合方式等)
- YOLOv5【网络结构】超详细解读总结！！！建议收藏✨✨！
- ❗重点参考：【目标检测】yolov5模型详解
YOLOv5（You Only Look Once version 5）是一个流行的目标检测算法，它是 YOLO 系列算法的最新版本之一。YOLOv5 以其高速和高精度的特点，在实时目标检测任务中表现出色。本次构建的是基于YOLOv5l的backbone（有修改），大致如下：
可以看出YOLOv5s包含输入、Backbone、Neck、Head、输出等几个主要部分
-
backbone:主干网络，其大多时候指的是提取特征的网络。主干网络的作用就是提取图片中的信息，供后面的网络使用。在Yolov5中，常见的Backbone网络包括CSP-Darknet53或ResNet。这些网络都是相对轻量级的，能够在保证较高检测精度的同时，尽可能地减少计算量和内存占用。
其结构主要有Conv模块、C3模块、SPPF模块。Conv模块主要由卷积层、BN层和激活函数组成，C3模块则将前面的特征图进行自适应聚合，SPPF模块通过全局特征与局部特征的加权融合，获取更全面的空间信息 -
neck:负责对Backbone提取的特征进行多尺度特征融合，并把这些特征传递给预测层。
例如，在Yolov5采用的PANet结构中，通过多次上采样、拼接、点和点积来设计聚合策略，以此更好地利用多尺度特征。 -
head:主要负责进行最终的回归预测，即利用Backbone骨干网络提取的特征图来检测目标的位置和类别。
最后，输出端是模型预测的结果，包括每个目标的类别和其对应的边界框坐标等信息。 -
组件：
-
CBS：由Conv+BN+SiLU激活函数组成；
-
SPPF：spatial pyramid pooling-fast
-
C3:3指3个卷积
Concat：张量拼接，会扩充两个张量的维度，例如26×26×256和26×26×512两个张量拼接，结果是26×26×768。
add：张量相加，张量直接相加，不会扩充维度，例如104×104×128和104×104×128相加，结果还是104×104×128。
Input
YOLOv5在输入端Input采用了Mosaic数据增强，参考了CutMix数据增强的方法，Mosaic数据增强由原来的两张图像提高到四张图像进行拼接，并对图像进行随机缩放，随机裁剪和随机排列。使用数据增强可以改善数据集中，小、中、大目标数据不均衡的问题。
主要步骤为：
- Mosaic
- Copy paste
- Random affine（Scale, Translation and Shear)
- Mixup
- Albumentations
- Augment HSV(Hue, Saturation, Value)
- Random horizontal flip.
采用Mosaic数据增强的方式有几个优点：
- 丰富数据集：随机使用4张图像，随机缩放后随机拼接，增加很多小目标，大大丰富了数据集，提高了网络的鲁棒性。
- 减少GPU占用：随机拼接的方式让一张图像可以计算四张图像的数据，减少每个batch的数量，即使只有一个GPU，也能得到较好的结果。
- 同时通过对识别物体的裁剪，使模型根据局部特征识别物体，有助于被遮挡物体的检测，从而提升了模型的检测能力。
Backbone
- Conv模块（前文的CBS结构）
- BN层具有防止过拟合和加速收敛的作用；
- SiLu激活函数是Sigmoid 加权线性组合，SiLU 函数也称为 swish 函数。
公式：silu(x)=x∗σ(x), where σ(x) is the logistic sigmoid.Silu函数处处可导，且连续光滑。Silu并非一个单调的函数，最大的缺点是计算量大。
- C3模块
该模块是对残差特征进行学习的主要模块，结构分为两支：
- Bottleneck堆叠
- 仅经过一个基本卷积模块
最后两支进行concat操作
C3中的Bottleneck:
借鉴了ResNet的残差结构，如下：
- 其中一路先进行1×1卷积将特征图的通道数减小一半，从而减少计算量，再通过3 ×3卷积提取特征，并且将通道数加倍，其输入与输出的通道数是不发生改变的。
- 另外一路通过shortcut进行残差连接，与第一路的输出特征图相加，从而实现特征融合。
在YOLOv5中，Backbone中的Bottleneck都默认使shortcut为True，而在Head中的Bottleneck都不使用shortcut。
- SPPF模块
SPPF由SPP改进而来，SPP先通过一个标准卷积模块将输入通道减半，然后分别做kernel-size为5，9，13的max pooling（对于不同的核大小，padding是自适应的）。
对三次最大池化的结果与未进行池化操作的数据进行concat，最终合并后channel数是原来的2倍。
yolo的SPP借鉴了空间金字塔的思想，通过SPP模块实现了局部特征和全部特征。经过局部特征与全矩特征相融合后，丰富了特征图的表达能力，有利于待检测图像中目标大小差异较大的情况，对yolo这种复杂的多目标检测的精度有很大的提升。
SPPF（Spatial Pyramid Pooling - Fast ）使用3个5×5的最大池化，代替原来的5×5、9×9、13×13最大池化，多个小尺寸池化核级联代替SPP模块中单个大尺寸池化核，从而在保留原有功能，即融合不同感受野的特征图，丰富特征图的表达能力的情况下，进一步提高了运行速度。
Neck
在Neck部分，yolov5主要采用了PANet结构。
PANet在FPN（feature pyramid network）上提取网络内特征层次结构，FPN中顶部信息流需要通过骨干网络(Backbone)逐层地往下传递，由于层数相对较多，因此计算量比较大（a)。
PANet在FPN的基础上又引入了一个自底向上(Bottom-up)的路径。经过自顶向下(Top-down)的特征融合后，再进行自底向上(Bottom-up)的特征融合，这样底层的位置信息也能够传递到深层，从而增强多个尺度上的定位能力。
- FPN(Feature Pyramid Networks) 特征金字塔模型：https://blog.csdn.net/u012856866/article/details/130271655
Head
Head部分主要用于检测目标，分别输出20×20，40×40和80×80的特征图大小，对应的是32×32，16×16和8×8像素的目标。
2.搭建模型
import torch.nn.functional as F
def autopad(k, p=None): # kernel, padding
# Pad to 'same'
if p is None:
p = k // 2 if isinstance(k, int) else [x // 2 for x in k] # auto-pad
return p
class Conv(nn.Module):
# Standard convolution
def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups
super().__init__()
self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=False)
self.bn = nn.BatchNorm2d(c2)
self.act = nn.SiLU() if act is True else (act if isinstance(act, nn.Module) else nn.Identity())
def forward(self, x):
return self.act(self.bn(self.conv(x)))
class Bottleneck(nn.Module):
# Standard bottleneck
def __init__(self, c1, c2, shortcut=True, g=1, e=0.5): # ch_in, ch_out, shortcut, groups, expansion
super().__init__()
c_ = int(c2 * e) # hidden channels
self.cv1 = Conv(c1, c_, 1, 1)
self.cv2 = Conv(c_, c2, 3, 1, g=g)
self.add = shortcut and c1 == c2
def forward(self, x):
return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))
class C3(nn.Module):
# CSP Bottleneck with 3 convolutions
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
super().__init__()
c_ = int(c2 * e) # hidden channels
self.cv1 = Conv(c1, c_, 1, 1)
self.cv2 = Conv(c1, c_, 1, 1)
self.cv3 = Conv(2 * c_, c2, 1) # act=FReLU(c2)
self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)))
def forward(self, x):
return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), dim=1))
class SPPF(nn.Module):
# Spatial Pyramid Pooling - Fast (SPPF) layer for YOLOv5 by Glenn Jocher
def __init__(self, c1, c2, k=5): # equivalent to SPP(k=(5, 9, 13))
super().__init__()
c_ = c1 // 2 # hidden channels
self.cv1 = Conv(c1, c_, 1, 1)
self.cv2 = Conv(c_ * 4, c2, 1, 1)
self.m = nn.MaxPool2d(kernel_size=k, stride=1, padding=k // 2)
def forward(self, x):
x = self.cv1(x)
with warnings.catch_warnings():
warnings.simplefilter('ignore') # suppress torch 1.9.0 max_pool2d() warning
y1 = self.m(x)
y2 = self.m(y1)
return self.cv2(torch.cat([x, y1, y2, self.m(y2)], 1))
"""
这个是YOLOv5, 6.0版本的主干网络，这里进行复现
（注：有部分删改，详细讲解将在后续进行展开）
"""
class YOLOv5_backbone(nn.Module):
def __init__(self):
super(YOLOv5_backbone, self).__init__()
self.Conv_1 = Conv(3, 64, 3, 2, 2)
self.Conv_2 = Conv(64, 128, 3, 2)
self.C3_3 = C3(128,128)
self.Conv_4 = Conv(128, 256, 3, 2)
self.C3_5 = C3(256,256)
self.Conv_6 = Conv(256, 512, 3, 2)
self.C3_7 = C3(512, 512)
self.Conv_8 = Conv(512, 1024, 3, 2)
self.C3_9 = C3(1024, 1024)
self.SPPF = SPPF(1024, 1024, 5)
# 全连接网络层，用于分类
self.classifier = nn.Sequential(
nn.Linear(in_features=65536, out_features=100),
nn.ReLU(),
nn.Dropout(0.2),
nn.Linear(in_features=100, out_features=4)
)
def forward(self, x):
x = self.Conv_1(x)
x = self.Conv_2(x)
x = self.C3_3(x)
x = self.Conv_4(x)
x = self.C3_5(x)
x = self.Conv_6(x)
x = self.C3_7(x)
x = self.Conv_8(x)
x = self.C3_9(x)
x = self.SPPF(x)
x = torch.flatten(x, start_dim=1)
x = self.classifier(x)
return x
model = YOLOv5_backbone().to(device)
model
YOLOv5_backbone(
(Conv_1): Conv(
(conv): Conv2d(3, 64, kernel_size=(3, 3), stride=(2, 2), padding=(2, 2), bias=False)
(bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
(Conv_2): Conv(
(conv): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
(C3_3): C3(
(cv1): Conv(
(conv): Conv2d(128, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(128, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv3): Conv(
(conv): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
(m): Sequential(
(0): Bottleneck(
(cv1): Conv(
(conv): Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
)
)
)
(Conv_4): Conv(
(conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
(C3_5): C3(
(cv1): Conv(
(conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv3): Conv(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
(m): Sequential(
(0): Bottleneck(
(cv1): Conv(
(conv): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
)
)
)
(Conv_6): Conv(
(conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
(C3_7): C3(
(cv1): Conv(
(conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv3): Conv(
(conv): Conv2d(512, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
(m): Sequential(
(0): Bottleneck(
(cv1): Conv(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
)
)
)
(Conv_8): Conv(
(conv): Conv2d(512, 1024, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
(C3_9): C3(
(cv1): Conv(
(conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv3): Conv(
(conv): Conv2d(1024, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
(m): Sequential(
(0): Bottleneck(
(cv1): Conv(
(conv): Conv2d(512, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
)
)
)
(SPPF): SPPF(
(cv1): Conv(
(conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
(cv2): Conv(
(conv): Conv2d(2048, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(act): SiLU()
)
(m): MaxPool2d(kernel_size=5, stride=1, padding=2, dilation=1, ceil_mode=False)
)
(classifier): Sequential(
(0): Linear(in_features=65536, out_features=100, bias=True)
(1): ReLU()
(2): Dropout(p=0.2, inplace=False)
(3): Linear(in_features=100, out_features=4, bias=True)
)
)
import torchsummary as ts
ts.summary(model,(3,224,224))
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [-1, 64, 113, 113] 1,728
BatchNorm2d-2 [-1, 64, 113, 113] 128
SiLU-3 [-1, 64, 113, 113] 0
Conv-4 [-1, 64, 113, 113] 0
Conv2d-5 [-1, 128, 57, 57] 73,728
BatchNorm2d-6 [-1, 128, 57, 57] 256
SiLU-7 [-1, 128, 57, 57] 0
Conv-8 [-1, 128, 57, 57] 0
Conv2d-9 [-1, 64, 57, 57] 8,192
BatchNorm2d-10 [-1, 64, 57, 57] 128
SiLU-11 [-1, 64, 57, 57] 0
Conv-12 [-1, 64, 57, 57] 0
Conv2d-13 [-1, 64, 57, 57] 4,096
BatchNorm2d-14 [-1, 64, 57, 57] 128
SiLU-15 [-1, 64, 57, 57] 0
Conv-16 [-1, 64, 57, 57] 0
Conv2d-17 [-1, 64, 57, 57] 36,864
BatchNorm2d-18 [-1, 64, 57, 57] 128
SiLU-19 [-1, 64, 57, 57] 0
Conv-20 [-1, 64, 57, 57] 0
Bottleneck-21 [-1, 64, 57, 57] 0
Conv2d-22 [-1, 64, 57, 57] 8,192
BatchNorm2d-23 [-1, 64, 57, 57] 128
SiLU-24 [-1, 64, 57, 57] 0
Conv-25 [-1, 64, 57, 57] 0
Conv2d-26 [-1, 128, 57, 57] 16,384
BatchNorm2d-27 [-1, 128, 57, 57] 256
SiLU-28 [-1, 128, 57, 57] 0
Conv-29 [-1, 128, 57, 57] 0
C3-30 [-1, 128, 57, 57] 0
Conv2d-31 [-1, 256, 29, 29] 294,912
BatchNorm2d-32 [-1, 256, 29, 29] 512
SiLU-33 [-1, 256, 29, 29] 0
Conv-34 [-1, 256, 29, 29] 0
Conv2d-35 [-1, 128, 29, 29] 32,768
BatchNorm2d-36 [-1, 128, 29, 29] 256
SiLU-37 [-1, 128, 29, 29] 0
Conv-38 [-1, 128, 29, 29] 0
Conv2d-39 [-1, 128, 29, 29] 16,384
BatchNorm2d-40 [-1, 128, 29, 29] 256
SiLU-41 [-1, 128, 29, 29] 0
Conv-42 [-1, 128, 29, 29] 0
Conv2d-43 [-1, 128, 29, 29] 147,456
BatchNorm2d-44 [-1, 128, 29, 29] 256
SiLU-45 [-1, 128, 29, 29] 0
Conv-46 [-1, 128, 29, 29] 0
Bottleneck-47 [-1, 128, 29, 29] 0
Conv2d-48 [-1, 128, 29, 29] 32,768
BatchNorm2d-49 [-1, 128, 29, 29] 256
SiLU-50 [-1, 128, 29, 29] 0
Conv-51 [-1, 128, 29, 29] 0
Conv2d-52 [-1, 256, 29, 29] 65,536
BatchNorm2d-53 [-1, 256, 29, 29] 512
SiLU-54 [-1, 256, 29, 29] 0
Conv-55 [-1, 256, 29, 29] 0
C3-56 [-1, 256, 29, 29] 0
Conv2d-57 [-1, 512, 15, 15] 1,179,648
BatchNorm2d-58 [-1, 512, 15, 15] 1,024
SiLU-59 [-1, 512, 15, 15] 0
Conv-60 [-1, 512, 15, 15] 0
Conv2d-61 [-1, 256, 15, 15] 131,072
BatchNorm2d-62 [-1, 256, 15, 15] 512
SiLU-63 [-1, 256, 15, 15] 0
Conv-64 [-1, 256, 15, 15] 0
Conv2d-65 [-1, 256, 15, 15] 65,536
BatchNorm2d-66 [-1, 256, 15, 15] 512
SiLU-67 [-1, 256, 15, 15] 0
Conv-68 [-1, 256, 15, 15] 0
Conv2d-69 [-1, 256, 15, 15] 589,824
BatchNorm2d-70 [-1, 256, 15, 15] 512
SiLU-71 [-1, 256, 15, 15] 0
Conv-72 [-1, 256, 15, 15] 0
Bottleneck-73 [-1, 256, 15, 15] 0
Conv2d-74 [-1, 256, 15, 15] 131,072
BatchNorm2d-75 [-1, 256, 15, 15] 512
SiLU-76 [-1, 256, 15, 15] 0
Conv-77 [-1, 256, 15, 15] 0
Conv2d-78 [-1, 512, 15, 15] 262,144
BatchNorm2d-79 [-1, 512, 15, 15] 1,024
SiLU-80 [-1, 512, 15, 15] 0
Conv-81 [-1, 512, 15, 15] 0
C3-82 [-1, 512, 15, 15] 0
Conv2d-83 [-1, 1024, 8, 8] 4,718,592
BatchNorm2d-84 [-1, 1024, 8, 8] 2,048
SiLU-85 [-1, 1024, 8, 8] 0
Conv-86 [-1, 1024, 8, 8] 0
Conv2d-87 [-1, 512, 8, 8] 524,288
BatchNorm2d-88 [-1, 512, 8, 8] 1,024
SiLU-89 [-1, 512, 8, 8] 0
Conv-90 [-1, 512, 8, 8] 0
Conv2d-91 [-1, 512, 8, 8] 262,144
BatchNorm2d-92 [-1, 512, 8, 8] 1,024
SiLU-93 [-1, 512, 8, 8] 0
Conv-94 [-1, 512, 8, 8] 0
Conv2d-95 [-1, 512, 8, 8] 2,359,296
BatchNorm2d-96 [-1, 512, 8, 8] 1,024
SiLU-97 [-1, 512, 8, 8] 0
Conv-98 [-1, 512, 8, 8] 0
Bottleneck-99 [-1, 512, 8, 8] 0
Conv2d-100 [-1, 512, 8, 8] 524,288
BatchNorm2d-101 [-1, 512, 8, 8] 1,024
SiLU-102 [-1, 512, 8, 8] 0
Conv-103 [-1, 512, 8, 8] 0
Conv2d-104 [-1, 1024, 8, 8] 1,048,576
BatchNorm2d-105 [-1, 1024, 8, 8] 2,048
SiLU-106 [-1, 1024, 8, 8] 0
Conv-107 [-1, 1024, 8, 8] 0
C3-108 [-1, 1024, 8, 8] 0
Conv2d-109 [-1, 512, 8, 8] 524,288
BatchNorm2d-110 [-1, 512, 8, 8] 1,024
SiLU-111 [-1, 512, 8, 8] 0
Conv-112 [-1, 512, 8, 8] 0
MaxPool2d-113 [-1, 512, 8, 8] 0
MaxPool2d-114 [-1, 512, 8, 8] 0
MaxPool2d-115 [-1, 512, 8, 8] 0
Conv2d-116 [-1, 1024, 8, 8] 2,097,152
BatchNorm2d-117 [-1, 1024, 8, 8] 2,048
SiLU-118 [-1, 1024, 8, 8] 0
Conv-119 [-1, 1024, 8, 8] 0
SPPF-120 [-1, 1024, 8, 8] 0
Linear-121 [-1, 100] 6,553,700
ReLU-122 [-1, 100] 0
Dropout-123 [-1, 100] 0
Linear-124 [-1, 4] 404
================================================================
Total params: 21,729,592
Trainable params: 21,729,592
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.57
Forward/backward pass size (MB): 137.59
Params size (MB): 82.89
Estimated Total Size (MB): 221.06
----------------------------------------------------------------
三、训练模型
1.编写训练函数
# 训练循环
def train(dataloader, model, loss_fn, optimizer):
size = len(dataloader.dataset) # 训练集的大小
num_batches = len(dataloader) # 批次数目, (size/batch_size，向上取整)
train_loss, train_acc = 0, 0 # 初始化训练损失和正确率
for X, y in dataloader: # 获取图片及其标签
X, y = X.to(device), y.to(device)
# 计算预测误差
pred = model(X) # 网络输出
loss = loss_fn(pred, y) # 计算网络输出和真实值之间的差距，targets为真实值，计算二者差值即为损失
# 反向传播
optimizer.zero_grad() # grad属性归零
loss.backward() # 反向传播
optimizer.step() # 每一步自动更新
# 记录acc与loss
train_acc += (pred.argmax(1) == y).type(torch.float).sum().item()
train_loss += loss.item()
train_acc /= size
train_loss /= num_batches
return train_acc, train_loss
2.编写训练函数
def test (dataloader, model, loss_fn):
size = len(dataloader.dataset) # 测试集的大小
num_batches = len(dataloader) # 批次数目, (size/batch_size，向上取整)
test_loss, test_acc = 0, 0
# 当不进行训练时，停止梯度更新，节省计算内存消耗
with torch.no_grad():
for imgs, target in dataloader:
imgs, target = imgs.to(device), target.to(device)
# 计算loss
target_pred = model(imgs)
loss = loss_fn(target_pred, target)
test_loss += loss.item()
test_acc += (target_pred.argmax(1) == target).type(torch.float).sum().item()
test_acc /= size
test_loss /= num_batches
return test_acc, test_loss
3.设置学习率
# 调用官方动态学习率接口时使用
learn_rate = 1e-4 # 初始学习率
lambda1 = lambda epoch: 0.94 ** (epoch // 4)
optimizer = torch.optim.Adam(model.parameters(), lr=learn_rate)
scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lambda1) #选定调整方法
3.正式训练
import copy
loss_fn = nn.CrossEntropyLoss() # 创建损失函数
#optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)
epochs = 50
train_loss = []
train_acc = []
test_loss = []
test_acc = []
best_acc = 0 # 设置一个最佳准确率，作为最佳模型的判别指标
for epoch in range(epochs):
model.train()
epoch_train_acc, epoch_train_loss = train(train_dl, model, loss_fn, optimizer)
scheduler.step()
model.eval()
epoch_test_acc, epoch_test_loss = test(test_dl, model, loss_fn)
# 保存最佳模型到 best_model
if epoch_test_acc > best_acc:
best_acc = epoch_test_acc
best_model = copy.deepcopy(model)
train_acc.append(epoch_train_acc)
train_loss.append(epoch_train_loss)
test_acc.append(epoch_test_acc)
test_loss.append(epoch_test_loss)
# 获取当前的学习率
lr = optimizer.state_dict()['param_groups'][0]['lr']
template = ('Epoch:{:2d}, Train_acc:{:.1f}%, Train_loss:{:.3f}, Test_acc:{:.1f}%, Test_loss:{:.3f}, Lr:{:.2E}')
print(template.format(epoch+1, epoch_train_acc*100, epoch_train_loss,
epoch_test_acc*100, epoch_test_loss, lr))
# 保存最佳模型到文件中
PATH = './P9_best_model.pth' # 保存的参数文件名
torch.save(model.state_dict(), PATH)
print('Done')
Epoch: 1, Train_acc:57.2%, Train_loss:1.093, Test_acc:68.0%, Test_loss:0.585, Lr:1.00E-04
Epoch: 2, Train_acc:66.6%, Train_loss:0.815, Test_acc:64.9%, Test_loss:1.037, Lr:1.00E-04
Epoch: 3, Train_acc:73.3%, Train_loss:0.757, Test_acc:84.4%, Test_loss:0.377, Lr:1.00E-04
... ...
Epoch:38, Train_acc:98.7%, Train_loss:0.043, Test_acc:93.8%, Test_loss:0.297, Lr:5.73E-05
Epoch:39, Train_acc:98.3%, Train_loss:0.053, Test_acc:95.6%, Test_loss:0.198, Lr:5.73E-05
Epoch:40, Train_acc:99.0%, Train_loss:0.025, Test_acc:93.8%, Test_loss:0.585, Lr:5.39E-05
Epoch:41, Train_acc:100.0%, Train_loss:0.006, Test_acc:94.2%, Test_loss:0.328, Lr:5.39E-05
Epoch:42, Train_acc:99.2%, Train_loss:0.027, Test_acc:96.0%, Test_loss:0.195, Lr:5.39E-05
Epoch:43, Train_acc:99.7%, Train_loss:0.010, Test_acc:94.7%, Test_loss:0.250, Lr:5.39E-05
Epoch:44, Train_acc:100.0%, Train_loss:0.003, Test_acc:95.6%, Test_loss:0.212, Lr:5.06E-05
Epoch:45, Train_acc:100.0%, Train_loss:0.002, Test_acc:95.6%, Test_loss:0.239, Lr:5.06E-05
Epoch:46, Train_acc:99.7%, Train_loss:0.008, Test_acc:96.0%, Test_loss:0.217, Lr:5.06E-05
Epoch:47, Train_acc:99.7%, Train_loss:0.015, Test_acc:96.0%, Test_loss:0.190, Lr:5.06E-05
Epoch:48, Train_acc:97.9%, Train_loss:0.068, Test_acc:93.3%, Test_loss:0.516, Lr:4.76E-05
Epoch:49, Train_acc:97.7%, Train_loss:0.074, Test_acc:93.3%, Test_loss:0.286, Lr:4.76E-05
Epoch:50, Train_acc:99.1%, Train_loss:0.032, Test_acc:94.2%, Test_loss:0.270, Lr:4.76E-05
Done
四、结果可视化
1.Loss与Accuracy图
import matplotlib.pyplot as plt
#隐藏警告
import warnings
warnings.filterwarnings("ignore") #忽略警告信息
# plt.rcParams['font.sans-serif'] = ['SimHei'] # 用来正常显示中文标签
plt.rcParams['axes.unicode_minus'] = False # 用来正常显示负号
plt.rcParams['figure.dpi'] = 100 #分辨率
epochs_range = range(epochs)
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 and Validation 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 and Validation Loss')
plt.show()
from PIL import Image
classes = list(total_data.class_to_idx)
def predict_one_image(image_path, model, transform, classes):
test_img = Image.open(image_path).convert('RGB')
plt.imshow(test_img) # 展示预测的图片
test_img = transform(test_img)
img = test_img.to(device).unsqueeze(0)
model.eval()
output = model(img)
_,pred = torch.max(output,1)
pred_class = classes[pred]
print(f'预测结果是：{pred_class}')
predict_one_image(image_path='./weather_photos/sunrise/sunrise56.jpg',
model=model,
transform=train_transforms,
classes=classes)
预测结果是：sunrise
2.模型评估
best_model.eval()
epoch_test_acc, epoch_test_loss = test(test_dl, best_model, loss_fn)
epoch_test_acc, epoch_test_loss
(0.96, 0.1952975117749404)
五、总结
- 本周学习了YOLOv5l网络的Backbone模块但对于YOLOv5的其他模块仍旧模糊；
- 通过设置batch size为8，设置动态学习率和dropout层进一步巩固对深度学习训练的理解；
- 在50 epoch内达到96%的test acc.
扫一扫
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
