---

---

环境

• 系统: Linux
• 语言: Python3.8.10
• 深度学习框架: Pytorch2.0.0+cu118

步骤

环境设置

包引用

import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader, random_split
from torchvision import datasets, transforms

import pathlib, random, copy
from PIL import Image

import numpy as np
import matplotlib.pyplot as plt
from torchinfo import summary

声明一个全局的设备

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

数据准备

收集数据集信息

# 图像数据的路径
image_path = 'weather_photos'

# 所有图像的列表
image_list = list(pathlib.Path(image_path).glob('*/*'))

# 随机打印几张图像的信息
for _ in range(5):
	image = random.choice(image_list)
	print(f"{str(image)}, shape is: {np.array(Image.open(str(image))).shape}")

# 查看随机的20张图像
plt.figure(figsize=(20, 4))
for i in range(20):
	plt.subplot(2, 10, i+1)
	plt.axis('off')
	image = random.choice(image_list)
	plt.title(image.parts[-2])
	plt.imshow(Image.open(str(image)))

通过图像信息的获取可以发现图像的尺寸并不一致，因此需要在构建数据集的时候对图像做一些伸缩处理。

构建数据集

img_transform = transforms.Compose([
	transforms.Resize([224, 224]),
	transforms.ToTensor(),
	transforms.Normalize(
        mean=[0.485, 0.456, 0.406], 
        std=[0.229, 0.224, 0.225]
	)
])
dataset = datasets.ImageFolder(image_path, transform=img_transform)

在数据集中读取分类名称

class_names = [k for k in dataset.class_to_idx]
print(class_names)

划分训练、测试数据集

train_size = int(len(dataset) * 0.8)
test_size = len(dataset) - train_size
train_dataset, test_dataset = random_split(dataset, [train_size, test_size])

数据集划分批次

batch_size = 32
train_loader = DataLoader(train_dataset, shuffle=True, batch_size=batch_size)
test_loader = DataLoader(test_dataset, batch_size=batch_size)

模型设计

编写维持卷积前后图像大小不变的padding计算函数

def pad(kernel_size, padding=None):
	if padding is None:
		return kernel_size // 2 if isinstance(kernel_size, int) else [item //2 for item in kernel_size]
	return padding

YOLOv5_95">编写YOLOv5中使用的卷积模块

class Conv(nn.Module):
	def __init__(self, ch_in, ch_out, kernel_size, stride=1, padding=None, groups=1, activation=True):
		super().__init__()
		
		self.conv = nn.Conv2d(ch_in, ch_out, kernel_size, stride, pad(kernel_size, padding), groups=groups, bias=False)
		self.bn = nn.BatchNorm2d(ch_out)
		self.act = nn.SiLU() if activation is True else (activation if isinstance(activation, nn.Module) else nn.Identity())

	def forward(self, x):
		x = self.act(self.bn(self.conv(x)))
		return x

YOLOv5Bottleneck_110">编写YOLOv5中使用的Bottleneck模块

class Bottleneck(nn.Module):
	def __init__(self, ch_in, ch_out, shortcut=True, groups=1, factor=0.5):
		super().__init__()

		hidden_size = int(ch_out*factor)
		self.conv1 = Conv(ch_in, hidden_size, 1)
		self.conv2 = Conv(hidden_size, ch_out, 3)
		self.add = shortcut and ch_in == ch_out

	def forward(self, x):
		return x + self.conv2(self.conv1(x)) if self.add else self.conv2(self.conv1(x))

YOLOv5C3_125">编写YOLOv5中使用的C3模块

class C3(nn.Module):
	def __init__(self, ch_in, ch_out, n=1, shortcut=True, groups=1, factor=0.5):
		super().__init__()

		hidden_size = int(ch_out*factor)
		self.conv1 = Conv(ch_in, hidden_size, 1)
		self.conv2 = Conv(ch_in, hidden_size, 1)
		self.conv3 = Conv(2*hidden_size, ch_out, 1)
		self.m = nn.Sequential(*(Bottleneck(hidden_size, hidden_size) for _ in range(n)))

	def forward(self, x):
		return self.conv3(torch.cat((self.conv1(x), self.m(self.conv2(x))), dim=1))

YOLOv5SPPF_140">编写YOLOv5中使用SPPF模块

class SPPF(nn.Module):
	def __init__(self, ch_in, ch_out, kernel_size=5):
		super().__init__()

		hidden_size = ch_in // 2
		self.conv1 = Conv(ch_in, hidden_size, 1)
		self.conv2 = Conv(4*hidden_size, ch_out, 1)
		self.m = nn.MaxPool2d(kernel_size=kernel_size, stride=1, padding=kernel_size//2)

	def forward(self, x):
		x = self.conv1(x)
		y1 = self.m(x)
		y2 = self.m(y1)
		y3 = self.m(y2)
		return self.conv2(torch.cat([x, y1, y2, y3], dim=1))

基于以上模块编写本任务需要的网络结构

class Network(nn.Module):
	def __init__(self, num_classes):
		super().__init__()

		self.conv1 = Conv(3, 64, 3, 2, 2)
		self.conv2 = Conv(64, 128, 3, 2)
		self.c3_1 = C3(128, 128)
		self.conv3 = Conv(128, 256, 3, 2)
		self.c3_2 = C3(256, 256)
		self.conv4 = Conv(256, 512, 3, 2)
		self.c3_3 = C3(512, 512)
		self.conv5 = Conv(512, 1024, 3, 2)
		self.c3_4 = C3(1024, 1024)
		self.sppf = SPPF(1024, 1024, 5)

		self.classifier = nn.Sequential(
			nn.Linear(65536, 100),
			nn.ReLU(),
			nn.Linear(100, num_classes)
		)
	
	def forward(self, x):
		x = self.conv1(x)
		x = self.conv2(x)
		x = self.c3_1(x)
		x = self.conv3(x)
		x = self.c3_2(x)
		x = self.conv4(x)
		x = self.c3_3(x)
		x = self.conv5(x)
		x = self.c3_4(x)
		x = self.sppf(x)
		x = x.view(x.size(0), -1)
		x = self.classifier(x)
		return x
model = Network(len(class_names)).to(device)
print(model)
summary(model, input_size=(32, 3, 224, 224))

直接打印出的模型结构如下:

Network(
  (conv1): 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()
  )
  (conv2): 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_1): C3(
    (conv1): 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()
    )
    (conv2): 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()
    )
    (conv3): 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(
        (conv1): Conv(
          (conv): Conv2d(64, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)
          (bn): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
          (act): SiLU()
        )
        (conv2): Conv(
          (conv): Conv2d(32, 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()
        )
      )
    )
  )
  (conv3): 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_2): C3(
    (conv1): 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()
    )
    (conv2): 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()
    )
    (conv3): 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(
        (conv1): 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()
        )
        (conv2): Conv(
          (conv): Conv2d(64, 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()
        )
      )
    )
  )
  (conv4): 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_3): C3(
    (conv1): 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()
    )
    (conv2): 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()
    )
    (conv3): 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(
        (conv1): 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()
        )
        (conv2): Conv(
          (conv): Conv2d(128, 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()
        )
      )
    )
  )
  (conv5): 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_4): C3(
    (conv1): 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()
    )
    (conv2): 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()
    )
    (conv3): 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(
        (conv1): 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()
        )
        (conv2): Conv(
          (conv): Conv2d(256, 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(
    (conv1): 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()
    )
    (conv2): 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): Linear(in_features=100, out_features=4, bias=True)
  )
)

torchinfo库中的summary函数打印的结果如下：

===============================================================================================
Layer (type:depth-idx)                        Output Shape              Param #
===============================================================================================
Network                                       [32, 4]                   --
├─Conv: 1-1                                   [32, 64, 113, 113]        --
│    └─Conv2d: 2-1                            [32, 64, 113, 113]        1,728
│    └─BatchNorm2d: 2-2                       [32, 64, 113, 113]        128
│    └─SiLU: 2-3                              [32, 64, 113, 113]        --
├─Conv: 1-2                                   [32, 128, 57, 57]         --
│    └─Conv2d: 2-4                            [32, 128, 57, 57]         73,728
│    └─BatchNorm2d: 2-5                       [32, 128, 57, 57]         256
│    └─SiLU: 2-6                              [32, 128, 57, 57]         --
├─C3: 1-3                                     [32, 128, 57, 57]         --
│    └─Conv: 2-7                              [32, 64, 57, 57]          --
│    │    └─Conv2d: 3-1                       [32, 64, 57, 57]          8,192
│    │    └─BatchNorm2d: 3-2                  [32, 64, 57, 57]          128
│    │    └─SiLU: 3-3                         [32, 64, 57, 57]          --
│    └─Conv: 2-8                              [32, 64, 57, 57]          --
│    │    └─Conv2d: 3-4                       [32, 64, 57, 57]          8,192
│    │    └─BatchNorm2d: 3-5                  [32, 64, 57, 57]          128
│    │    └─SiLU: 3-6                         [32, 64, 57, 57]          --
│    └─Sequential: 2-9                        [32, 64, 57, 57]          --
│    │    └─Bottleneck: 3-7                   [32, 64, 57, 57]          20,672
│    └─Conv: 2-10                             [32, 128, 57, 57]         --
│    │    └─Conv2d: 3-8                       [32, 128, 57, 57]         16,384
│    │    └─BatchNorm2d: 3-9                  [32, 128, 57, 57]         256
│    │    └─SiLU: 3-10                        [32, 128, 57, 57]         --
├─Conv: 1-4                                   [32, 256, 29, 29]         --
│    └─Conv2d: 2-11                           [32, 256, 29, 29]         294,912
│    └─BatchNorm2d: 2-12                      [32, 256, 29, 29]         512
│    └─SiLU: 2-13                             [32, 256, 29, 29]         --
├─C3: 1-5                                     [32, 256, 29, 29]         --
│    └─Conv: 2-14                             [32, 128, 29, 29]         --
│    │    └─Conv2d: 3-11                      [32, 128, 29, 29]         32,768
│    │    └─BatchNorm2d: 3-12                 [32, 128, 29, 29]         256
│    │    └─SiLU: 3-13                        [32, 128, 29, 29]         --
│    └─Conv: 2-15                             [32, 128, 29, 29]         --
│    │    └─Conv2d: 3-14                      [32, 128, 29, 29]         32,768
│    │    └─BatchNorm2d: 3-15                 [32, 128, 29, 29]         256
│    │    └─SiLU: 3-16                        [32, 128, 29, 29]         --
│    └─Sequential: 2-16                       [32, 128, 29, 29]         --
│    │    └─Bottleneck: 3-17                  [32, 128, 29, 29]         82,304
│    └─Conv: 2-17                             [32, 256, 29, 29]         --
│    │    └─Conv2d: 3-18                      [32, 256, 29, 29]         65,536
│    │    └─BatchNorm2d: 3-19                 [32, 256, 29, 29]         512
│    │    └─SiLU: 3-20                        [32, 256, 29, 29]         --
├─Conv: 1-6                                   [32, 512, 15, 15]         --
│    └─Conv2d: 2-18                           [32, 512, 15, 15]         1,179,648
│    └─BatchNorm2d: 2-19                      [32, 512, 15, 15]         1,024
│    └─SiLU: 2-20                             [32, 512, 15, 15]         --
├─C3: 1-7                                     [32, 512, 15, 15]         --
│    └─Conv: 2-21                             [32, 256, 15, 15]         --
│    │    └─Conv2d: 3-21                      [32, 256, 15, 15]         131,072
│    │    └─BatchNorm2d: 3-22                 [32, 256, 15, 15]         512
│    │    └─SiLU: 3-23                        [32, 256, 15, 15]         --
│    └─Conv: 2-22                             [32, 256, 15, 15]         --
│    │    └─Conv2d: 3-24                      [32, 256, 15, 15]         131,072
│    │    └─BatchNorm2d: 3-25                 [32, 256, 15, 15]         512
│    │    └─SiLU: 3-26                        [32, 256, 15, 15]         --
│    └─Sequential: 2-23                       [32, 256, 15, 15]         --
│    │    └─Bottleneck: 3-27                  [32, 256, 15, 15]         328,448
│    └─Conv: 2-24                             [32, 512, 15, 15]         --
│    │    └─Conv2d: 3-28                      [32, 512, 15, 15]         262,144
│    │    └─BatchNorm2d: 3-29                 [32, 512, 15, 15]         1,024
│    │    └─SiLU: 3-30                        [32, 512, 15, 15]         --
├─Conv: 1-8                                   [32, 1024, 8, 8]          --
│    └─Conv2d: 2-25                           [32, 1024, 8, 8]          4,718,592
│    └─BatchNorm2d: 2-26                      [32, 1024, 8, 8]          2,048
│    └─SiLU: 2-27                             [32, 1024, 8, 8]          --
├─C3: 1-9                                     [32, 1024, 8, 8]          --
│    └─Conv: 2-28                             [32, 512, 8, 8]           --
│    │    └─Conv2d: 3-31                      [32, 512, 8, 8]           524,288
│    │    └─BatchNorm2d: 3-32                 [32, 512, 8, 8]           1,024
│    │    └─SiLU: 3-33                        [32, 512, 8, 8]           --
│    └─Conv: 2-29                             [32, 512, 8, 8]           --
│    │    └─Conv2d: 3-34                      [32, 512, 8, 8]           524,288
│    │    └─BatchNorm2d: 3-35                 [32, 512, 8, 8]           1,024
│    │    └─SiLU: 3-36                        [32, 512, 8, 8]           --
│    └─Sequential: 2-30                       [32, 512, 8, 8]           --
│    │    └─Bottleneck: 3-37                  [32, 512, 8, 8]           1,312,256
│    └─Conv: 2-31                             [32, 1024, 8, 8]          --
│    │    └─Conv2d: 3-38                      [32, 1024, 8, 8]          1,048,576
│    │    └─BatchNorm2d: 3-39                 [32, 1024, 8, 8]          2,048
│    │    └─SiLU: 3-40                        [32, 1024, 8, 8]          --
├─SPPF: 1-10                                  [32, 1024, 8, 8]          --
│    └─Conv: 2-32                             [32, 512, 8, 8]           --
│    │    └─Conv2d: 3-41                      [32, 512, 8, 8]           524,288
│    │    └─BatchNorm2d: 3-42                 [32, 512, 8, 8]           1,024
│    │    └─SiLU: 3-43                        [32, 512, 8, 8]           --
│    └─MaxPool2d: 2-33                        [32, 512, 8, 8]           --
│    └─MaxPool2d: 2-34                        [32, 512, 8, 8]           --
│    └─MaxPool2d: 2-35                        [32, 512, 8, 8]           --
│    └─Conv: 2-36                             [32, 1024, 8, 8]          --
│    │    └─Conv2d: 3-44                      [32, 1024, 8, 8]          2,097,152
│    │    └─BatchNorm2d: 3-45                 [32, 1024, 8, 8]          2,048
│    │    └─SiLU: 3-46                        [32, 1024, 8, 8]          --
├─Sequential: 1-11                            [32, 4]                   --
│    └─Linear: 2-37                           [32, 100]                 6,553,700
│    └─ReLU: 2-38                             [32, 100]                 --
│    └─Linear: 2-39                           [32, 4]                   404
===============================================================================================
Total params: 19,987,832
Trainable params: 19,987,832
Non-trainable params: 0
Total mult-adds (G): 64.43
===============================================================================================
Input size (MB): 19.27
Forward/backward pass size (MB): 2027.63
Params size (MB): 79.95
Estimated Total Size (MB): 2126.85
===============================================================================================

模型训练

编写训练函数

def train(train_loader, model, loss_fn, optimizer):
	model.train()
	train_loss, train_acc = 0, 0
	num_batches = len(train_loader)
	size = len(train_loader.dataset)

	for x, y in train_loader:
		x, y = x.to(device), y.to(device)

		pred = model(x)
		loss = loss_fn(pred, y)

		optimizer.zero_grad()
		loss.backward()
		optimizer.step()

		train_loss += loss.item()
		train_acc += (pred.argmax(1) == y).type(torch.float).sum().item()

	train_loss /= num_batches
	train_acc /= size

	return train_loss, train_acc

def test(test_loader, model, loss_fn):
	model.eval()
	test_loss, test_acc = 0, 0
	num_batches = len(test_loader)
	size = len(test_loader.dataset)

	with torch.no_grad():
		for x, y in test_loader:
			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_loss /= num_batches
	test_acc /= size

	return test_loss, test_acc

开始模型的训练

epochs = 60
loss_fn = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=1e-4)
best_acc = 0
best_model_path = 'best_p9_model.pth'

train_loss, train_acc = [], []
test_loss, test_acc = [], []

for epoch in range(epochs):
	epoch_train_loss, epoch_train_acc = train(train_loader, model, loss_fn, optimizer)
	epoch_test_loss, epoch_test_acc = test(test_loader, model, loss_fn)

	if best_acc < epoch_test_acc:
		best_acc = epoch_test_acc
		best_model = copy.deepcopy(model)

	train_loss.append(epoch_train_loss)
	train_acc.append(epoch_train_acc)
	test_loss.append(epoch_test_loss)
	test_acc.append(epoch_test_acc)

	lr = optimizer.state_dict()['param_groups'][0]['lr']

	print(f"Epoch: {epoch+1}, TrainLoss: {epoch_train_loss:.3f}, TrainAcc: {epoch_train_acc*100:.1f},TestLoss: {epoch_test_loss:.3f}, TestAcc: {epoch_test_acc*100:.1f}, learning_rate: {lr}")
print(f"training finished, save best model to : {best_model_path})")
torch.save(best_model.state_dict(), best_model_path)
print("done")

训练过程打印日志如下

Epoch: 1, TrainLoss: 0.986, TrainAcc: 57.2,TestLoss: 2.137, TestAcc: 25.3, learning_rate: 0.0001
Epoch: 2, TrainLoss: 0.725, TrainAcc: 76.2,TestLoss: 0.486, TestAcc: 87.6, learning_rate: 0.0001
Epoch: 3, TrainLoss: 0.368, TrainAcc: 84.7,TestLoss: 0.310, TestAcc: 87.6, learning_rate: 0.0001
Epoch: 4, TrainLoss: 0.295, TrainAcc: 89.9,TestLoss: 0.329, TestAcc: 90.7, learning_rate: 0.0001
Epoch: 5, TrainLoss: 0.407, TrainAcc: 87.2,TestLoss: 0.288, TestAcc: 88.9, learning_rate: 0.0001
Epoch: 6, TrainLoss: 0.316, TrainAcc: 89.4,TestLoss: 0.354, TestAcc: 89.8, learning_rate: 0.0001
Epoch: 7, TrainLoss: 0.347, TrainAcc: 92.1,TestLoss: 0.244, TestAcc: 92.9, learning_rate: 0.0001
Epoch: 8, TrainLoss: 0.206, TrainAcc: 93.1,TestLoss: 0.313, TestAcc: 94.2, learning_rate: 0.0001
Epoch: 9, TrainLoss: 0.204, TrainAcc: 92.4,TestLoss: 0.227, TestAcc: 90.2, learning_rate: 0.0001
Epoch: 10, TrainLoss: 0.151, TrainAcc: 95.4,TestLoss: 0.242, TestAcc: 92.9, learning_rate: 0.0001
Epoch: 11, TrainLoss: 0.146, TrainAcc: 95.6,TestLoss: 0.314, TestAcc: 88.9, learning_rate: 0.0001
Epoch: 12, TrainLoss: 0.223, TrainAcc: 91.7,TestLoss: 0.769, TestAcc: 92.0, learning_rate: 0.0001
Epoch: 13, TrainLoss: 0.155, TrainAcc: 95.2,TestLoss: 0.223, TestAcc: 92.4, learning_rate: 0.0001
Epoch: 14, TrainLoss: 0.267, TrainAcc: 93.9,TestLoss: 0.280, TestAcc: 93.3, learning_rate: 0.0001
Epoch: 15, TrainLoss: 0.194, TrainAcc: 93.3,TestLoss: 0.345, TestAcc: 89.3, learning_rate: 0.0001
Epoch: 16, TrainLoss: 0.283, TrainAcc: 91.3,TestLoss: 0.267, TestAcc: 92.4, learning_rate: 0.0001
Epoch: 17, TrainLoss: 0.183, TrainAcc: 94.3,TestLoss: 1.779, TestAcc: 84.4, learning_rate: 0.0001
Epoch: 18, TrainLoss: 0.161, TrainAcc: 95.7,TestLoss: 0.279, TestAcc: 90.7, learning_rate: 0.0001
Epoch: 19, TrainLoss: 0.100, TrainAcc: 95.7,TestLoss: 0.249, TestAcc: 93.8, learning_rate: 0.0001
Epoch: 20, TrainLoss: 0.134, TrainAcc: 97.0,TestLoss: 0.252, TestAcc: 91.1, learning_rate: 0.0001
Epoch: 21, TrainLoss: 0.236, TrainAcc: 94.0,TestLoss: 0.264, TestAcc: 88.0, learning_rate: 0.0001
Epoch: 22, TrainLoss: 0.199, TrainAcc: 93.1,TestLoss: 0.251, TestAcc: 94.7, learning_rate: 0.0001
Epoch: 23, TrainLoss: 0.243, TrainAcc: 95.2,TestLoss: 0.425, TestAcc: 88.0, learning_rate: 0.0001
Epoch: 24, TrainLoss: 0.181, TrainAcc: 94.8,TestLoss: 0.390, TestAcc: 86.7, learning_rate: 0.0001
Epoch: 25, TrainLoss: 0.138, TrainAcc: 97.4,TestLoss: 0.337, TestAcc: 91.1, learning_rate: 0.0001
Epoch: 26, TrainLoss: 0.212, TrainAcc: 96.6,TestLoss: 0.358, TestAcc: 90.2, learning_rate: 0.0001
Epoch: 27, TrainLoss: 0.289, TrainAcc: 92.4,TestLoss: 0.239, TestAcc: 94.2, learning_rate: 0.0001
Epoch: 28, TrainLoss: 0.220, TrainAcc: 95.6,TestLoss: 0.280, TestAcc: 88.4, learning_rate: 0.0001
Epoch: 29, TrainLoss: 0.177, TrainAcc: 95.6,TestLoss: 0.216, TestAcc: 92.9, learning_rate: 0.0001
Epoch: 30, TrainLoss: 0.116, TrainAcc: 96.3,TestLoss: 0.240, TestAcc: 92.0, learning_rate: 0.0001
Epoch: 31, TrainLoss: 0.065, TrainAcc: 98.0,TestLoss: 0.230, TestAcc: 92.4, learning_rate: 0.0001
Epoch: 32, TrainLoss: 0.097, TrainAcc: 98.0,TestLoss: 0.261, TestAcc: 92.9, learning_rate: 0.0001
Epoch: 33, TrainLoss: 0.084, TrainAcc: 97.9,TestLoss: 0.262, TestAcc: 92.0, learning_rate: 0.0001
Epoch: 34, TrainLoss: 0.113, TrainAcc: 96.2,TestLoss: 0.257, TestAcc: 95.1, learning_rate: 0.0001
Epoch: 35, TrainLoss: 0.071, TrainAcc: 97.8,TestLoss: 0.284, TestAcc: 92.0, learning_rate: 0.0001
Epoch: 36, TrainLoss: 0.238, TrainAcc: 95.2,TestLoss: 0.210, TestAcc: 92.0, learning_rate: 0.0001
Epoch: 37, TrainLoss: 0.175, TrainAcc: 96.9,TestLoss: 0.259, TestAcc: 92.9, learning_rate: 0.0001
Epoch: 38, TrainLoss: 0.129, TrainAcc: 95.8,TestLoss: 0.315, TestAcc: 92.0, learning_rate: 0.0001
Epoch: 39, TrainLoss: 0.077, TrainAcc: 98.0,TestLoss: 0.233, TestAcc: 91.6, learning_rate: 0.0001
Epoch: 40, TrainLoss: 0.092, TrainAcc: 97.3,TestLoss: 0.266, TestAcc: 89.3, learning_rate: 0.0001
Epoch: 41, TrainLoss: 0.064, TrainAcc: 98.0,TestLoss: 0.248, TestAcc: 92.0, learning_rate: 0.0001
Epoch: 42, TrainLoss: 0.062, TrainAcc: 99.2,TestLoss: 0.211, TestAcc: 93.3, learning_rate: 0.0001
Epoch: 43, TrainLoss: 0.098, TrainAcc: 97.2,TestLoss: 0.359, TestAcc: 90.7, learning_rate: 0.0001
Epoch: 44, TrainLoss: 0.153, TrainAcc: 97.0,TestLoss: 0.411, TestAcc: 89.3, learning_rate: 0.0001
Epoch: 45, TrainLoss: 0.234, TrainAcc: 96.9,TestLoss: 0.198, TestAcc: 92.0, learning_rate: 0.0001
Epoch: 46, TrainLoss: 0.042, TrainAcc: 98.6,TestLoss: 0.191, TestAcc: 93.3, learning_rate: 0.0001
Epoch: 47, TrainLoss: 0.033, TrainAcc: 98.9,TestLoss: 0.141, TestAcc: 96.9, learning_rate: 0.0001
Epoch: 48, TrainLoss: 0.012, TrainAcc: 99.7,TestLoss: 0.202, TestAcc: 94.7, learning_rate: 0.0001
Epoch: 49, TrainLoss: 0.016, TrainAcc: 99.1,TestLoss: 0.171, TestAcc: 93.8, learning_rate: 0.0001
Epoch: 50, TrainLoss: 0.011, TrainAcc: 99.6,TestLoss: 0.274, TestAcc: 93.8, learning_rate: 0.0001
Epoch: 51, TrainLoss: 0.014, TrainAcc: 99.8,TestLoss: 0.233, TestAcc: 94.2, learning_rate: 0.0001
Epoch: 52, TrainLoss: 0.263, TrainAcc: 98.7,TestLoss: 0.233, TestAcc: 91.6, learning_rate: 0.0001
Epoch: 53, TrainLoss: 0.284, TrainAcc: 92.7,TestLoss: 0.680, TestAcc: 92.9, learning_rate: 0.0001
Epoch: 54, TrainLoss: 0.334, TrainAcc: 90.9,TestLoss: 0.332, TestAcc: 91.1, learning_rate: 0.0001
Epoch: 55, TrainLoss: 0.261, TrainAcc: 94.4,TestLoss: 0.498, TestAcc: 90.7, learning_rate: 0.0001
Epoch: 56, TrainLoss: 0.144, TrainAcc: 95.9,TestLoss: 0.376, TestAcc: 88.4, learning_rate: 0.0001
Epoch: 57, TrainLoss: 0.080, TrainAcc: 97.3,TestLoss: 0.296, TestAcc: 92.4, learning_rate: 0.0001
Epoch: 58, TrainLoss: 0.033, TrainAcc: 99.2,TestLoss: 0.226, TestAcc: 93.3, learning_rate: 0.0001
Epoch: 59, TrainLoss: 0.023, TrainAcc: 99.0,TestLoss: 0.327, TestAcc: 93.8, learning_rate: 0.0001
Epoch: 60, TrainLoss: 0.073, TrainAcc: 98.0,TestLoss: 0.347, TestAcc: 90.7, learning_rate: 0.0001
training finished, save best model to : best_p9_model.pth)
done

训练过程图表展示

epoch_ranges = range(epochs)

plt.figure(figsize=(20, 4))
plt.subplot(121)
plt.plot(epoch_ranges, train_loss, label='train loss')
plt.plot(epoch_ranges, test_loss, label='validation loss')
plt.legend(loc='upper right')
plt.title('Loss')

plt.subplot(122)
plt.plot(epoch_ranges, train_acc, label='train accuracy')
plt.plot(epoch_ranges, test_acc, label='validation accuracy')
plt.legend(loc='lower right')
plt.title('Accuracy')

模型效果展示

载入最佳模型

model.load_state_dict(torch.load(best_model_path))
model.to(device)

编写预测函数

def predict(model, image_path):
	image = Image.open(image_path)
	image = img_transform(image)
	image = image.unsqueeze(0).to(device)
	model.eval()
	with torch.no_grad():
		pred = model(image)
	return class_names[pred.argmax(1).item()]

执行预测并展示

评估应当只用测试集中的数据，这里没有再使用数据集进行反归一化，会有一部分训练集中的数据参与其中，导致预测的结果非常好（可能是假象）

plt.figure(figsize=(20, 4))
for i in range(20):
	plt.subplot(2, 10, i+1)
	image = random.choice(image_list)
	real_label = image.parts[-2]
	pred_label = predict(model, str(image))
	plt.title(f"R:{real_label}, P:{pred_label}")
	plt.axis('off')
	plt.imshow(Image.open(str(image)))

总结与心得体会

• YOLOv5的骨干网络中大量使用了1x1卷积，只用来将特征图重新映射到不同通道的特征图中，执行效率比执行大核卷积快
• 骨干网络中特征图大小的缩减并没用像普通的卷积网络一样使用池化层，而是使用卷积通过调整stride和padding属性来实现，这样做会比直接使用池化层多一些可训练参数，可能会增加一些模型的拟合能力
• 通过本次任务学习到了跳跃连接应该怎样编写代码，还有SPPF模块的实现
• 模型的评估应该不包含训练集中的数据，不然展示的结果并不真实