- 🍨 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 | 接辅导、项目定制
目录
- 0. 总结
- 1. 设置GPU
- 2. 导入数据及处理部分
- 3. 划分数据集
- 4. 模型构建部分
- 5. 设置超参数：定义损失函数，学习率，以及根据学习率定义优化器等
- 6. 训练函数
- 7. 测试函数
- 8. 正式训练
- 9. 结果可视化
- 10. 模型的保存
- 11. 使用训练好的模型进行预测
0. 总结
数据导入及处理部分：本次数据导入没有使用torchvision自带的数据集，需要将原始数据进行处理包括数据导入，查看数据分类情况，定义transforms，进行数据类型转换等操作。
划分数据集：划定训练集测试集后，再使用torch.utils.data中的DataLoader()分别加载上一步处理好的训练及测试数据，查看批处理维度.
模型构建部分：resnet-50v2
设置超参数：在这之前需要定义损失函数，学习率（动态学习率），以及根据学习率定义优化器（例如SGD随机梯度下降），用来在训练中更新参数，最小化损失函数。
定义训练函数：函数的传入的参数有四个，分别是设置好的DataLoader(),定义好的模型，损失函数，优化器。函数内部初始化损失准确率为0，接着开始循环，使用DataLoader()获取一个批次的数据，对这个批次的数据带入模型得到预测值，然后使用损失函数计算得到损失值。接下来就是进行反向传播以及使用优化器优化参数，梯度清零放在反向传播之前或者是使用优化器优化之后都是可以的，一般是默认放在反向传播之前。
定义测试函数：函数传入的参数相比训练函数少了优化器，只需传入设置好的DataLoader(),定义好的模型，损失函数。此外除了处理批次数据时无需再设置梯度清零、返向传播以及优化器优化参数，其余部分均和训练函数保持一致。
训练过程：定义训练次数，有几次就使用整个数据集进行几次训练，初始化四个空list分别存储每次训练及测试的准确率及损失。使用model.train()开启训练模式，调用训练函数得到准确率及损失。使用model.eval()将模型设置为评估模式，调用测试函数得到准确率及损失。接着就是将得到的训练及测试的准确率及损失存储到相应list中并合并打印出来，得到每一次整体训练后的准确率及损失。
结果可视化
模型的保存，调取及使用。在PyTorch中，通常使用 torch.save(model.state_dict(), ‘model.pth’) 保存模型的参数，使用 model.load_state_dict(torch.load(‘model.pth’)) 加载参数。
需要改进优化的地方：确保模型和数据的一致性，都存到GPU或者CPU;注意numclasses不要直接用默认的1000，需要根据实际数据集改进；实例化模型也要注意numclasses这个参数；此外注意测试模型需要用（3,224,224）3表示通道数，这和tensorflow定义的顺序是不用的（224,224,3），做代码转换时需要注意。
import torch
import torch.nn as nn
import torchvision
from torchvision import datasets,transforms
from torch.utils.data import DataLoader
import torchvision.models as models
import torch.nn.functional as F
import os,PIL,pathlib
import matplotlib.pyplot as plt
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
1. 设置GPU
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
device
device(type='cuda')
2. 导入数据及处理部分
# 获取数据分布情况
path_dir = './data/bird_photos/'
path_dir = pathlib.Path(path_dir)
paths = list(path_dir.glob('*'))
# classNames = [str(path).split("\\")[-1] for path in paths] # ['Bananaquit', 'Black Skimmer', 'Black Throated Bushtiti', 'Cockatoo']
classNames = [path.parts[-1] for path in paths]
classNames
['Bananaquit', 'Black Skimmer', 'Black Throated Bushtiti', 'Cockatoo']
# 定义transforms 并处理数据
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 装换为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], # 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。
std = [0.229,0.224,0.225]
)
])
test_transforms = transforms.Compose([
transforms.Resize([224,224]),
transforms.ToTensor(),
transforms.Normalize(
mean = [0.485,0.456,0.406],
std = [0.229,0.224,0.225]
)
])
total_data = datasets.ImageFolder('./data/bird_photos/',transform = train_transforms)
total_data
Dataset ImageFolder
Number of datapoints: 565
Root location: ./data/bird_photos/
StandardTransform
Transform: Compose(
Resize(size=[224, 224], interpolation=bilinear, max_size=None, antialias=True)
RandomHorizontalFlip(p=0.5)
ToTensor()
Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
)
total_data.class_to_idx
{'Bananaquit': 0,
'Black Skimmer': 1,
'Black Throated Bushtiti': 2,
'Cockatoo': 3}
3. 划分数据集
# 划分数据集
train_size = int(len(total_data) * 0.8)
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 0x1d5973216c0>,
<torch.utils.data.dataset.Subset at 0x1d5973201c0>)
# 定义DataLoader用于数据集的加载
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
)
# 观察数据维度
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([32, 3, 224, 224])
Shape of y: torch.Size([32]) torch.int64
4. 模型构建部分
import torch
import torch.nn as nn
import torchvision.models as models
import torchvision.transforms as transforms
class Block2(nn.Module):
def __init__(self, in_channels, filters, stride=1, conv_shortcut=False):
super(Block2, self).__init__()
self.conv_shortcut = conv_shortcut
self.stride = stride
self.preact_bn = nn.BatchNorm2d(in_channels)
self.preact_relu = nn.ReLU(inplace=True)
if self.conv_shortcut:
self.shortcut = nn.Conv2d(in_channels, 4 * filters, kernel_size=1, stride=stride)
else:
self.shortcut = None if stride == 1 else nn.MaxPool2d(kernel_size=1, stride=stride)
self.conv1 = nn.Conv2d(in_channels, filters, kernel_size=1, stride=1, bias=False)
self.bn1 = nn.BatchNorm2d(filters)
self.relu1 = nn.ReLU(inplace=True)
self.conv2 = nn.Conv2d(filters, filters, kernel_size=3, stride=stride, padding=1, bias=False)
self.bn2 = nn.BatchNorm2d(filters)
self.relu2 = nn.ReLU(inplace=True)
self.conv3 = nn.Conv2d(filters, 4 * filters, kernel_size=1)
def forward(self, x):
preact = self.preact_bn(x)
preact = self.preact_relu(preact)
if self.conv_shortcut:
shortcut = self.shortcut(preact)
else:
shortcut = x if self.stride == 1 else self.shortcut(x)
x = self.conv1(preact)
x = self.bn1(x)
x = self.relu1(x)
x = self.conv2(x)
x = self.bn2(x)
x = self.relu2(x)
x = self.conv3(x)
x += shortcut
return x
class Stack2(nn.Module):
def __init__(self, in_channels, filters, blocks, stride1=2):
super(Stack2, self).__init__()
self.blocks = nn.ModuleList()
self.blocks.append(Block2(in_channels, filters, stride=stride1, conv_shortcut=True))
for _ in range(1, blocks - 1):
self.blocks.append(Block2(4 * filters, filters))
self.blocks.append(Block2(4 * filters, filters, stride=1))
def forward(self, x):
for block in self.blocks:
x = block(x)
return x
class ResNet50V2(nn.Module):
def __init__(self, num_classes=1000):
super(ResNet50V2, self).__init__()
self.conv1_pad = nn.ZeroPad2d(padding=(3, 3, 3, 3))
self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, bias=False)
self.conv1_bn = nn.BatchNorm2d(64)
self.conv1_relu = nn.ReLU(inplace=True)
self.pool1_pad = nn.ZeroPad2d(padding=(1, 1, 1, 1))
self.pool1 = nn.MaxPool2d(3, stride=2)
self.stack1 = Stack2(64, 64, 3)
self.stack2 = Stack2(256, 128, 4)
self.stack3 = Stack2(512, 256, 6)
self.stack4 = Stack2(1024, 512, 3)
self.post_bn = nn.BatchNorm2d(2048)
self.post_relu = nn.ReLU(inplace=True)
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.fc = nn.Linear(2048, num_classes)
def forward(self, x):
x = self.conv1_pad(x)
x = self.conv1(x)
x = self.conv1_bn(x)
x = self.conv1_relu(x)
x = self.pool1_pad(x)
x = self.pool1(x)
x = self.stack1(x)
x = self.stack2(x)
x = self.stack3(x)
x = self.stack4(x)
x = self.post_bn(x)
x = self.post_relu(x)
x = self.avgpool(x)
x = torch.flatten(x, 1)
x = self.fc(x)
return x
# Now, instantiate and use the model
model = ResNet50V2(num_classes=len(classNames))
model.to(device)
ResNet50V2(
(conv1_pad): ZeroPad2d((3, 3, 3, 3))
(conv1): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), bias=False)
(conv1_bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv1_relu): ReLU(inplace=True)
(pool1_pad): ZeroPad2d((1, 1, 1, 1))
(pool1): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)
(stack1): Stack2(
(blocks): ModuleList(
(0): Block2(
(preact_bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(preact_relu): ReLU(inplace=True)
(shortcut): Conv2d(64, 256, kernel_size=(1, 1), stride=(2, 2))
(conv1): Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu1): ReLU(inplace=True)
(conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu2): ReLU(inplace=True)
(conv3): Conv2d(64, 256, kernel_size=(1, 1), stride=(1, 1))
)
(1-2): 2 x Block2(
(preact_bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(preact_relu): ReLU(inplace=True)
(conv1): Conv2d(256, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu1): ReLU(inplace=True)
(conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu2): ReLU(inplace=True)
(conv3): Conv2d(64, 256, kernel_size=(1, 1), stride=(1, 1))
)
)
)
(stack2): Stack2(
(blocks): ModuleList(
(0): Block2(
(preact_bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(preact_relu): ReLU(inplace=True)
(shortcut): Conv2d(256, 512, kernel_size=(1, 1), stride=(2, 2))
(conv1): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu1): ReLU(inplace=True)
(conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu2): ReLU(inplace=True)
(conv3): Conv2d(128, 512, kernel_size=(1, 1), stride=(1, 1))
)
(1-3): 3 x Block2(
(preact_bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(preact_relu): ReLU(inplace=True)
(conv1): Conv2d(512, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu1): ReLU(inplace=True)
(conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu2): ReLU(inplace=True)
(conv3): Conv2d(128, 512, kernel_size=(1, 1), stride=(1, 1))
)
)
)
(stack3): Stack2(
(blocks): ModuleList(
(0): Block2(
(preact_bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(preact_relu): ReLU(inplace=True)
(shortcut): Conv2d(512, 1024, kernel_size=(1, 1), stride=(2, 2))
(conv1): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu1): ReLU(inplace=True)
(conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu2): ReLU(inplace=True)
(conv3): Conv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1))
)
(1-5): 5 x Block2(
(preact_bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(preact_relu): ReLU(inplace=True)
(conv1): Conv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu1): ReLU(inplace=True)
(conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu2): ReLU(inplace=True)
(conv3): Conv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1))
)
)
)
(stack4): Stack2(
(blocks): ModuleList(
(0): Block2(
(preact_bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(preact_relu): ReLU(inplace=True)
(shortcut): Conv2d(1024, 2048, kernel_size=(1, 1), stride=(2, 2))
(conv1): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu1): ReLU(inplace=True)
(conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu2): ReLU(inplace=True)
(conv3): Conv2d(512, 2048, kernel_size=(1, 1), stride=(1, 1))
)
(1-2): 2 x Block2(
(preact_bn): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(preact_relu): ReLU(inplace=True)
(conv1): Conv2d(2048, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu1): ReLU(inplace=True)
(conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu2): ReLU(inplace=True)
(conv3): Conv2d(512, 2048, kernel_size=(1, 1), stride=(1, 1))
)
)
)
(post_bn): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(post_relu): ReLU(inplace=True)
(avgpool): AdaptiveAvgPool2d(output_size=(1, 1))
(fc): Linear(in_features=2048, out_features=4, bias=True)
)
# 查看模型详情
import torchsummary as summary
summary.summary(model,(3,224,224))
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
ZeroPad2d-1 [-1, 3, 230, 230] 0
Conv2d-2 [-1, 64, 112, 112] 9,408
BatchNorm2d-3 [-1, 64, 112, 112] 128
ReLU-4 [-1, 64, 112, 112] 0
ZeroPad2d-5 [-1, 64, 114, 114] 0
MaxPool2d-6 [-1, 64, 56, 56] 0
BatchNorm2d-7 [-1, 64, 56, 56] 128
ReLU-8 [-1, 64, 56, 56] 0
Conv2d-9 [-1, 256, 28, 28] 16,640
Conv2d-10 [-1, 64, 56, 56] 4,096
BatchNorm2d-11 [-1, 64, 56, 56] 128
ReLU-12 [-1, 64, 56, 56] 0
Conv2d-13 [-1, 64, 28, 28] 36,864
BatchNorm2d-14 [-1, 64, 28, 28] 128
ReLU-15 [-1, 64, 28, 28] 0
Conv2d-16 [-1, 256, 28, 28] 16,640
Block2-17 [-1, 256, 28, 28] 0
BatchNorm2d-18 [-1, 256, 28, 28] 512
ReLU-19 [-1, 256, 28, 28] 0
Conv2d-20 [-1, 64, 28, 28] 16,384
BatchNorm2d-21 [-1, 64, 28, 28] 128
ReLU-22 [-1, 64, 28, 28] 0
Conv2d-23 [-1, 64, 28, 28] 36,864
BatchNorm2d-24 [-1, 64, 28, 28] 128
ReLU-25 [-1, 64, 28, 28] 0
Conv2d-26 [-1, 256, 28, 28] 16,640
Block2-27 [-1, 256, 28, 28] 0
BatchNorm2d-28 [-1, 256, 28, 28] 512
ReLU-29 [-1, 256, 28, 28] 0
Conv2d-30 [-1, 64, 28, 28] 16,384
BatchNorm2d-31 [-1, 64, 28, 28] 128
ReLU-32 [-1, 64, 28, 28] 0
Conv2d-33 [-1, 64, 28, 28] 36,864
BatchNorm2d-34 [-1, 64, 28, 28] 128
ReLU-35 [-1, 64, 28, 28] 0
Conv2d-36 [-1, 256, 28, 28] 16,640
Block2-37 [-1, 256, 28, 28] 0
Stack2-38 [-1, 256, 28, 28] 0
BatchNorm2d-39 [-1, 256, 28, 28] 512
ReLU-40 [-1, 256, 28, 28] 0
Conv2d-41 [-1, 512, 14, 14] 131,584
Conv2d-42 [-1, 128, 28, 28] 32,768
BatchNorm2d-43 [-1, 128, 28, 28] 256
ReLU-44 [-1, 128, 28, 28] 0
Conv2d-45 [-1, 128, 14, 14] 147,456
BatchNorm2d-46 [-1, 128, 14, 14] 256
ReLU-47 [-1, 128, 14, 14] 0
Conv2d-48 [-1, 512, 14, 14] 66,048
Block2-49 [-1, 512, 14, 14] 0
BatchNorm2d-50 [-1, 512, 14, 14] 1,024
ReLU-51 [-1, 512, 14, 14] 0
Conv2d-52 [-1, 128, 14, 14] 65,536
BatchNorm2d-53 [-1, 128, 14, 14] 256
ReLU-54 [-1, 128, 14, 14] 0
Conv2d-55 [-1, 128, 14, 14] 147,456
BatchNorm2d-56 [-1, 128, 14, 14] 256
ReLU-57 [-1, 128, 14, 14] 0
Conv2d-58 [-1, 512, 14, 14] 66,048
Block2-59 [-1, 512, 14, 14] 0
BatchNorm2d-60 [-1, 512, 14, 14] 1,024
ReLU-61 [-1, 512, 14, 14] 0
Conv2d-62 [-1, 128, 14, 14] 65,536
BatchNorm2d-63 [-1, 128, 14, 14] 256
ReLU-64 [-1, 128, 14, 14] 0
Conv2d-65 [-1, 128, 14, 14] 147,456
BatchNorm2d-66 [-1, 128, 14, 14] 256
ReLU-67 [-1, 128, 14, 14] 0
Conv2d-68 [-1, 512, 14, 14] 66,048
Block2-69 [-1, 512, 14, 14] 0
BatchNorm2d-70 [-1, 512, 14, 14] 1,024
ReLU-71 [-1, 512, 14, 14] 0
Conv2d-72 [-1, 128, 14, 14] 65,536
BatchNorm2d-73 [-1, 128, 14, 14] 256
ReLU-74 [-1, 128, 14, 14] 0
Conv2d-75 [-1, 128, 14, 14] 147,456
BatchNorm2d-76 [-1, 128, 14, 14] 256
ReLU-77 [-1, 128, 14, 14] 0
Conv2d-78 [-1, 512, 14, 14] 66,048
Block2-79 [-1, 512, 14, 14] 0
Stack2-80 [-1, 512, 14, 14] 0
BatchNorm2d-81 [-1, 512, 14, 14] 1,024
ReLU-82 [-1, 512, 14, 14] 0
Conv2d-83 [-1, 1024, 7, 7] 525,312
Conv2d-84 [-1, 256, 14, 14] 131,072
BatchNorm2d-85 [-1, 256, 14, 14] 512
ReLU-86 [-1, 256, 14, 14] 0
Conv2d-87 [-1, 256, 7, 7] 589,824
BatchNorm2d-88 [-1, 256, 7, 7] 512
ReLU-89 [-1, 256, 7, 7] 0
Conv2d-90 [-1, 1024, 7, 7] 263,168
Block2-91 [-1, 1024, 7, 7] 0
BatchNorm2d-92 [-1, 1024, 7, 7] 2,048
ReLU-93 [-1, 1024, 7, 7] 0
Conv2d-94 [-1, 256, 7, 7] 262,144
BatchNorm2d-95 [-1, 256, 7, 7] 512
ReLU-96 [-1, 256, 7, 7] 0
Conv2d-97 [-1, 256, 7, 7] 589,824
BatchNorm2d-98 [-1, 256, 7, 7] 512
ReLU-99 [-1, 256, 7, 7] 0
Conv2d-100 [-1, 1024, 7, 7] 263,168
Block2-101 [-1, 1024, 7, 7] 0
BatchNorm2d-102 [-1, 1024, 7, 7] 2,048
ReLU-103 [-1, 1024, 7, 7] 0
Conv2d-104 [-1, 256, 7, 7] 262,144
BatchNorm2d-105 [-1, 256, 7, 7] 512
ReLU-106 [-1, 256, 7, 7] 0
Conv2d-107 [-1, 256, 7, 7] 589,824
BatchNorm2d-108 [-1, 256, 7, 7] 512
ReLU-109 [-1, 256, 7, 7] 0
Conv2d-110 [-1, 1024, 7, 7] 263,168
Block2-111 [-1, 1024, 7, 7] 0
BatchNorm2d-112 [-1, 1024, 7, 7] 2,048
ReLU-113 [-1, 1024, 7, 7] 0
Conv2d-114 [-1, 256, 7, 7] 262,144
BatchNorm2d-115 [-1, 256, 7, 7] 512
ReLU-116 [-1, 256, 7, 7] 0
Conv2d-117 [-1, 256, 7, 7] 589,824
BatchNorm2d-118 [-1, 256, 7, 7] 512
ReLU-119 [-1, 256, 7, 7] 0
Conv2d-120 [-1, 1024, 7, 7] 263,168
Block2-121 [-1, 1024, 7, 7] 0
BatchNorm2d-122 [-1, 1024, 7, 7] 2,048
ReLU-123 [-1, 1024, 7, 7] 0
Conv2d-124 [-1, 256, 7, 7] 262,144
BatchNorm2d-125 [-1, 256, 7, 7] 512
ReLU-126 [-1, 256, 7, 7] 0
Conv2d-127 [-1, 256, 7, 7] 589,824
BatchNorm2d-128 [-1, 256, 7, 7] 512
ReLU-129 [-1, 256, 7, 7] 0
Conv2d-130 [-1, 1024, 7, 7] 263,168
Block2-131 [-1, 1024, 7, 7] 0
BatchNorm2d-132 [-1, 1024, 7, 7] 2,048
ReLU-133 [-1, 1024, 7, 7] 0
Conv2d-134 [-1, 256, 7, 7] 262,144
BatchNorm2d-135 [-1, 256, 7, 7] 512
ReLU-136 [-1, 256, 7, 7] 0
Conv2d-137 [-1, 256, 7, 7] 589,824
BatchNorm2d-138 [-1, 256, 7, 7] 512
ReLU-139 [-1, 256, 7, 7] 0
Conv2d-140 [-1, 1024, 7, 7] 263,168
Block2-141 [-1, 1024, 7, 7] 0
Stack2-142 [-1, 1024, 7, 7] 0
BatchNorm2d-143 [-1, 1024, 7, 7] 2,048
ReLU-144 [-1, 1024, 7, 7] 0
Conv2d-145 [-1, 2048, 4, 4] 2,099,200
Conv2d-146 [-1, 512, 7, 7] 524,288
BatchNorm2d-147 [-1, 512, 7, 7] 1,024
ReLU-148 [-1, 512, 7, 7] 0
Conv2d-149 [-1, 512, 4, 4] 2,359,296
BatchNorm2d-150 [-1, 512, 4, 4] 1,024
ReLU-151 [-1, 512, 4, 4] 0
Conv2d-152 [-1, 2048, 4, 4] 1,050,624
Block2-153 [-1, 2048, 4, 4] 0
BatchNorm2d-154 [-1, 2048, 4, 4] 4,096
ReLU-155 [-1, 2048, 4, 4] 0
Conv2d-156 [-1, 512, 4, 4] 1,048,576
BatchNorm2d-157 [-1, 512, 4, 4] 1,024
ReLU-158 [-1, 512, 4, 4] 0
Conv2d-159 [-1, 512, 4, 4] 2,359,296
BatchNorm2d-160 [-1, 512, 4, 4] 1,024
ReLU-161 [-1, 512, 4, 4] 0
Conv2d-162 [-1, 2048, 4, 4] 1,050,624
Block2-163 [-1, 2048, 4, 4] 0
BatchNorm2d-164 [-1, 2048, 4, 4] 4,096
ReLU-165 [-1, 2048, 4, 4] 0
Conv2d-166 [-1, 512, 4, 4] 1,048,576
BatchNorm2d-167 [-1, 512, 4, 4] 1,024
ReLU-168 [-1, 512, 4, 4] 0
Conv2d-169 [-1, 512, 4, 4] 2,359,296
BatchNorm2d-170 [-1, 512, 4, 4] 1,024
ReLU-171 [-1, 512, 4, 4] 0
Conv2d-172 [-1, 2048, 4, 4] 1,050,624
Block2-173 [-1, 2048, 4, 4] 0
Stack2-174 [-1, 2048, 4, 4] 0
BatchNorm2d-175 [-1, 2048, 4, 4] 4,096
ReLU-176 [-1, 2048, 4, 4] 0
AdaptiveAvgPool2d-177 [-1, 2048, 1, 1] 0
Linear-178 [-1, 4] 8,196
================================================================
Total params: 23,527,620
Trainable params: 23,527,620
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.57
Forward/backward pass size (MB): 101.68
Params size (MB): 89.75
Estimated Total Size (MB): 192.01
----------------------------------------------------------------
5. 设置超参数：定义损失函数，学习率，以及根据学习率定义优化器等
# loss_fn = nn.CrossEntropyLoss() # 创建损失函数
# learn_rate = 1e-3 # 初始学习率
# def adjust_learning_rate(optimizer,epoch,start_lr):
# # 每两个epoch 衰减到原来的0.98
# lr = start_lr * (0.92 ** (epoch//2))
# for param_group in optimizer.param_groups:
# param_group['lr'] = lr
# optimizer = torch.optim.Adam(model.parameters(),lr=learn_rate)
# 调用官方接口示例
loss_fn = nn.CrossEntropyLoss()
learn_rate = 1e-4
lambda1 = lambda epoch:(0.92**(epoch//2))
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
6. 训练函数
# 训练函数
def train(dataloader,model,loss_fn,optimizer):
size = len(dataloader.dataset) # 训练集大小
num_batches = len(dataloader) # 批次数目
train_loss,train_acc = 0,0
for X,y in dataloader:
X,y = X.to(device),y.to(device)
# 计算预测误差
pred = model(X)
loss = loss_fn(pred,y)
# 反向传播
optimizer.zero_grad()
loss.backward()
optimizer.step()
# 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
7. 测试函数
# 测试函数
def test(dataloader,model,loss_fn):
size = len(dataloader.dataset)
num_batches = 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)
# 计算loss
pred = model(X)
loss = loss_fn(pred,y)
test_acc += (pred.argmax(1)==y).type(torch.float).sum().item()
test_loss += loss.item()
test_acc /= size
test_loss /= num_batches
return test_acc,test_loss
8. 正式训练
import copy
epochs = 40
train_acc = []
train_loss = []
test_acc = []
test_loss = []
best_acc = 0.0
for epoch in range(epochs):
# 更新学习率——使用自定义学习率时使用
# adjust_learning_rate(optimizer,epoch,learn_rate)
model.train()
epoch_train_acc,epoch_train_loss = train(train_dl,model,loss_fn,optimizer)
scheduler.step() # 更新学习率——调用官方动态学习率时使用
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))
print('Done')
Epoch: 1,Train_acc:48.0%,Train_loss:1.221,Test_acc:19.5%,Test_loss:1.435,Lr:1.00E-04
Epoch: 2,Train_acc:72.1%,Train_loss:0.746,Test_acc:36.3%,Test_loss:1.754,Lr:9.20E-05
Epoch: 3,Train_acc:85.2%,Train_loss:0.453,Test_acc:74.3%,Test_loss:0.690,Lr:9.20E-05
Epoch: 4,Train_acc:90.9%,Train_loss:0.288,Test_acc:71.7%,Test_loss:1.046,Lr:8.46E-05
Epoch: 5,Train_acc:93.1%,Train_loss:0.236,Test_acc:73.5%,Test_loss:1.107,Lr:8.46E-05
Epoch: 6,Train_acc:94.5%,Train_loss:0.162,Test_acc:72.6%,Test_loss:0.840,Lr:7.79E-05
Epoch: 7,Train_acc:96.5%,Train_loss:0.268,Test_acc:76.1%,Test_loss:0.703,Lr:7.79E-05
Epoch: 8,Train_acc:94.2%,Train_loss:0.234,Test_acc:78.8%,Test_loss:0.803,Lr:7.16E-05
Epoch: 9,Train_acc:94.5%,Train_loss:0.163,Test_acc:69.0%,Test_loss:1.384,Lr:7.16E-05
Epoch:10,Train_acc:94.9%,Train_loss:0.172,Test_acc:78.8%,Test_loss:0.606,Lr:6.59E-05
Epoch:11,Train_acc:96.7%,Train_loss:0.125,Test_acc:76.1%,Test_loss:0.757,Lr:6.59E-05
Epoch:12,Train_acc:97.6%,Train_loss:0.074,Test_acc:85.8%,Test_loss:0.452,Lr:6.06E-05
Epoch:13,Train_acc:97.8%,Train_loss:0.087,Test_acc:81.4%,Test_loss:0.592,Lr:6.06E-05
Epoch:14,Train_acc:98.0%,Train_loss:0.089,Test_acc:80.5%,Test_loss:0.617,Lr:5.58E-05
Epoch:15,Train_acc:95.4%,Train_loss:0.133,Test_acc:71.7%,Test_loss:1.433,Lr:5.58E-05
Epoch:16,Train_acc:97.6%,Train_loss:0.074,Test_acc:77.0%,Test_loss:0.772,Lr:5.13E-05
Epoch:17,Train_acc:98.5%,Train_loss:0.101,Test_acc:80.5%,Test_loss:0.843,Lr:5.13E-05
Epoch:18,Train_acc:97.8%,Train_loss:0.072,Test_acc:69.9%,Test_loss:1.233,Lr:4.72E-05
Epoch:19,Train_acc:98.5%,Train_loss:0.079,Test_acc:81.4%,Test_loss:0.866,Lr:4.72E-05
Epoch:20,Train_acc:97.6%,Train_loss:0.070,Test_acc:79.6%,Test_loss:0.767,Lr:4.34E-05
Epoch:21,Train_acc:98.0%,Train_loss:0.356,Test_acc:78.8%,Test_loss:0.836,Lr:4.34E-05
Epoch:22,Train_acc:96.2%,Train_loss:0.126,Test_acc:78.8%,Test_loss:0.697,Lr:4.00E-05
Epoch:23,Train_acc:99.1%,Train_loss:0.071,Test_acc:78.8%,Test_loss:0.655,Lr:4.00E-05
Epoch:24,Train_acc:97.8%,Train_loss:0.068,Test_acc:84.1%,Test_loss:0.527,Lr:3.68E-05
Epoch:25,Train_acc:98.2%,Train_loss:0.115,Test_acc:77.0%,Test_loss:0.790,Lr:3.68E-05
Epoch:26,Train_acc:98.0%,Train_loss:0.138,Test_acc:80.5%,Test_loss:0.657,Lr:3.38E-05
Epoch:27,Train_acc:98.0%,Train_loss:0.154,Test_acc:83.2%,Test_loss:0.536,Lr:3.38E-05
Epoch:28,Train_acc:98.9%,Train_loss:0.046,Test_acc:80.5%,Test_loss:0.576,Lr:3.11E-05
Epoch:29,Train_acc:98.7%,Train_loss:0.086,Test_acc:81.4%,Test_loss:0.569,Lr:3.11E-05
Epoch:30,Train_acc:99.8%,Train_loss:0.039,Test_acc:77.9%,Test_loss:0.906,Lr:2.86E-05
Epoch:31,Train_acc:99.1%,Train_loss:0.041,Test_acc:83.2%,Test_loss:0.521,Lr:2.86E-05
Epoch:32,Train_acc:99.3%,Train_loss:0.026,Test_acc:84.1%,Test_loss:0.510,Lr:2.63E-05
Epoch:33,Train_acc:99.8%,Train_loss:0.028,Test_acc:79.6%,Test_loss:0.566,Lr:2.63E-05
Epoch:34,Train_acc:99.8%,Train_loss:0.026,Test_acc:81.4%,Test_loss:0.553,Lr:2.42E-05
Epoch:35,Train_acc:98.5%,Train_loss:0.159,Test_acc:77.9%,Test_loss:0.684,Lr:2.42E-05
Epoch:36,Train_acc:99.3%,Train_loss:0.048,Test_acc:81.4%,Test_loss:0.591,Lr:2.23E-05
Epoch:37,Train_acc:99.3%,Train_loss:0.064,Test_acc:83.2%,Test_loss:0.509,Lr:2.23E-05
Epoch:38,Train_acc:99.1%,Train_loss:0.131,Test_acc:86.7%,Test_loss:0.597,Lr:2.05E-05
Epoch:39,Train_acc:99.1%,Train_loss:0.045,Test_acc:83.2%,Test_loss:0.652,Lr:2.05E-05
Epoch:40,Train_acc:99.8%,Train_loss:0.083,Test_acc:80.5%,Test_loss:0.627,Lr:1.89E-05
Done
9. 结果可视化
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 = 'Test Accuracy')
plt.plot(epochs_range,test_loss,label = 'Test Loss')
plt.legend(loc = 'lower right')
plt.title('Training and validation Loss')
plt.show()
10. 模型的保存
# 自定义模型保存
# 状态字典保存
torch.save(model.state_dict(),'./模型参数/J2_resnet50v2_model_state_dict.pth') # 仅保存状态字典
# 加载状态字典到模型
best_model = ResNet50V2(num_classes=len(classNames)).to(device) # 定义官方vgg16模型用来加载参数
best_model.load_state_dict(torch.load('./模型参数/J2_resnet50v2_model_state_dict.pth')) # 加载状态字典到模型
<All keys matched successfully>
11. 使用训练好的模型进行预测
# 指定路径图片预测
from PIL import Image
import torchvision.transforms as transforms
classes = list(total_data.class_to_idx) # 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) # 展示待预测的图片
test_img = transform(test_img)
img = test_img.to(device).unsqueeze(0)
model.eval()
output = model(img)
print(output) # 观察模型预测结果的输出数据
_,pred = torch.max(output,1)
pred_class = classes[pred]
print(f'预测结果是:{pred_class}')
# 预测训练集中的某张照片
predict_one_image(image_path='./data/bird_photos/Bananaquit/007.jpg',
model = model,
transform = test_transforms,
classes = classes
)
tensor([[ 8.8948, -4.9875, 1.8381, -6.7715]], device='cuda:0',
grad_fn=<AddmmBackward0>)
预测结果是:Bananaquit
classes
['Bananaquit', 'Black Skimmer', 'Black Throated Bushtiti', 'Cockatoo']
