Pytorch从零开始实战13

Pytorch从零开始实战——ResNet与DenseNet探索

本系列来源于365天深度学习训练营

原作者K同学

环境准备

本文基于Jupyter notebook，使用Python3.8，Pytorch2.0.1+cu118，torchvision0.15.2，需读者自行配置好环境且有一些深度学习理论基础。本次实验的目的是探索ResNet和DenseNet结合。
第一步，导入常用包

import torch
import torch.nn as nn
import matplotlib.pyplot as plt
import torchvision
import torchvision.transforms as transforms
import torchvision.datasets as datasets
import torch.nn.functional as F
import random
from time import time
import numpy as np
import pandas as pd
import datetime
import gc
import os
import copy
import warnings
os.environ['KMP_DUPLICATE_LIB_OK']='True'  # 用于避免jupyter环境突然关闭
torch.backends.cudnn.benchmark=True  # 用于加速GPU运算的代码

设置随机数种子

torch.manual_seed(428)
torch.cuda.manual_seed(428)
torch.cuda.manual_seed_all(428)
random.seed(428)
np.random.seed(428)

检查设备对象

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
device, torch.cuda.device_count() # # (device(type='cuda'), 2)

数据集

本次数据集是使用鸟的图片，分别有四种类别的鸟，根据鸟的类别名称存放在不同的文件夹中。
使用pathlib查看类别

import pathlib
data_dir = './data/bird_photos/'
data_dir = pathlib.Path(data_dir) # 转成pathlib.Path对象
data_paths = list(data_dir.glob('*')) 
classNames = [str(path).split("/")[2] for path in data_paths]
classNames # ['Black Throated Bushtiti', 'Cockatoo', 'Black Skimmer', 'Bananaquit']

使用transforms对数据集进行统一处理，并且根据文件夹名映射对应标签

all_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]) # 标准化
])

total_data = datasets.ImageFolder("./data/bird_photos/", transform=all_transforms)
total_data.class_to_idx
 # {'Bananaquit': 0,
 # 'Black Skimmer': 1,
 # 'Black Throated Bushtiti': 2,
 # 'Cockatoo': 3}

随机查看5张图片

def plotsample(data):
    fig, axs = plt.subplots(1, 5, figsize=(10, 10)) #建立子图
    for i in range(5):
        num = random.randint(0, len(data) - 1) #首先选取随机数，随机选取五次
        #抽取数据中对应的图像对象，make_grid函数可将任意格式的图像的通道数升为3，而不改变图像原始的数据
        #而展示图像用的imshow函数最常见的输入格式也是3通道
        npimg = torchvision.utils.make_grid(data[num][0]).numpy()
        nplabel = data[num][1] #提取标签 
        #将图像由(3, weight, height)转化为(weight, height, 3)，并放入imshow函数中读取
        axs[i].imshow(np.transpose(npimg, (1, 2, 0))) 
        axs[i].set_title(nplabel) #给每个子图加上标签
        axs[i].axis("off") #消除每个子图的坐标轴

plotsample(total_data)

根据8比2划分数据集和测试集，并且利用DataLoader划分批次和随机打乱

train_size = int(0.8 * len(total_data))
test_size  = len(total_data) - train_size
train_ds, test_ds = torch.utils.data.random_split(total_data, [train_size, test_size])

batch_size = 32
train_dl = torch.utils.data.DataLoader(train_ds,
                                        batch_size=batch_size,
                                        shuffle=True,
                                      )
test_dl = torch.utils.data.DataLoader(test_ds,
                                        batch_size=batch_size,
                                        shuffle=True,
                                     )

len(train_dl.dataset), len(test_dl.dataset) # (452, 113)

模型选择

本次模型采用模型集成，也就是说将一个输入特征分别送到两个不同的模型中去学习，到最后将输出特征融合后进行分类，每个模型可能对数据集的不同方面有更强的学习能力，因此通过集成，可以期望获得更全面、更强大的特征表示。
具体而言，本次实验使用了一个 DenseNet 和一个 ResNet，它们在网络结构和特征提取方面有所不同。DenseNet 使用密集连接的结构，充分利用了每一层的特征，而 ResNet 使用残差连接，有助于解决梯度消失问题，使得网络更易训练。

先建立ResNet模型，首先实现Block块，首先进行预激活层，包括标准化和ReLu激活函数，接着进行shortcut操作，如果conv_shortcut为True，会使用一个1x1卷积层进行变换，否则，如果stride为1，则进行恒等映射，否则使用1x1的最大池化。随后经过三个卷积层。在forward方法中，输入 x 经过预激活层，然后进行三个卷积操作，最后将shortcut和经过卷积的结果相加。这种结构使得梯度更容易反向传播，从而有助于训练深层网络。

class Block2(nn.Module):
    def __init__(self, in_channels, filters, kernel_size=3, stride=1, conv_shortcut=False):
        super(Block2, self).__init__()

        self.preact = nn.Sequential(
            nn.BatchNorm2d(in_channels),
            nn.ReLU()
        )

        if conv_shortcut:
            self.shortcut = nn.Conv2d(in_channels, 4 * filters, kernel_size=1, stride=stride)
        else:
            if stride == 1:
                self.shortcut = nn.Identity()
            else: 
                self.shortcut = nn.MaxPool2d(1, stride=stride)

        self.conv1 = nn.Sequential(
            nn.Conv2d(in_channels, filters, kernel_size=1, stride=1, bias=False),
            nn.BatchNorm2d(filters),
            nn.ReLU()
        )

        self.conv2 = nn.Sequential(
            nn.ZeroPad2d(padding=(1, 1, 1, 1)),
            nn.Conv2d(filters, filters, kernel_size=kernel_size, stride=stride, bias=False),
            nn.BatchNorm2d(filters),
            nn.ReLU()
        )

        self.conv3 = nn.Conv2d(filters, 4 * filters, kernel_size=1)

    def forward(self, x):
        preact = self.preact(x)
        shortcut = self.shortcut(preact)
        x = self.conv1(preact)
        x = self.conv2(x)
        x = self.conv3(x)
        out = shortcut + x
        return out

下面实现堆叠块，通过传入不同的参数去调用Block块，其中[Block2(4 * filters, filters) for i in range(0, blocks)]，使用 Python 中的列表解析创建了 blocks 个残差块。这些残差块的输入通道数为 4 * filters，以匹配前一个残差块的输出通道数。

class Stack2(nn.Module):
    def __init__(self, in_channels, filters, blocks, stride1=2):
        super(Stack2, self).__init__()

        self.blocks = nn.Sequential(
            Block2(in_channels, filters, conv_shortcut=True),
            *[Block2(4 * filters, filters) for i in range(0, blocks)],
            Block2(4 * filters, filters, stride=stride1)
        )

    def forward(self, x):
        return self.blocks(x)

下面实现ResNet网络主体，去掉全连接层。

class ResNetPart(nn.Module):
    def __init__(self, include_top=True, preact=True, num_classes=1000):
        super().__init__()

        self.conv1 = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False),
            nn.BatchNorm2d(64),
            nn.ReLU()
        )

        self.pool1 = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)

        self.conv2 = Stack2(64, 64, 3)
        self.conv3 = Stack2(256, 128, 4)
        self.conv4 = Stack2(512, 256, 6)
        self.conv5 = Stack2(1024, 512, 3, stride1=1)

        self.post = nn.Sequential(
            nn.BatchNorm2d(2048),
            nn.ReLU()
        )

        self.include_top = include_top
        if include_top:
            self.avg_pool = nn.AdaptiveAvgPool2d(1)

    def forward(self, x):
        x = self.conv1(x)
        x = self.pool1(x)
        x = self.conv2(x)
        x = self.conv3(x)
        x = self.conv4(x)
        x = self.conv5(x)
        x = self.post(x)

        if self.include_top:
            x = self.avg_pool(x)
            x = torch.flatten(x, 1)

        return x

下面实现DenseNet部分。首先对DenseLayer类定义，本次实验使用add_module函数，默认是用于向类中添加一个子模块，第一个参数为模块名，第二个参数为模块实例，其实相当于加到父类的nn.Sequential里面，所以调用的时候使用super().forward(x)，这段的核心是将输入 x 与新特征 t 进行通道维度上的连接，完成密集连接。

class DenseLayer(nn.Sequential):
    def __init__(self, num_input_features, growth_rate, bn_size, drop_rate):
        super().__init__()
        self.add_module("norm1", nn.BatchNorm2d(num_input_features))
        self.add_module("relu1", nn.ReLU(inplace=True))
        self.add_module("conv1", nn.Conv2d(num_input_features, bn_size * growth_rate, kernel_size=1, stride=1, bias=False))
        self.add_module("norm2", nn.BatchNorm2d(bn_size * growth_rate))
        self.add_module("relu2", nn.ReLU(inplace=True))
        self.add_module("conv2", nn.Conv2d(bn_size*growth_rate, growth_rate, kernel_size=3, stride=1, padding=1, bias=False))
        self.drop_rate = drop_rate

    def forward(self, x):
        t = super().forward(x)
        if self.drop_rate > 0:
            t = F.dropout(t, p=self.drop_rate, training=self.training)
        return torch.cat([x, t], 1)

下面是DenseBlock的实现，通过循环创建了多个DenseLayer。其中的 num_input_features + i * growth_rate 用于指定输入通道的数量，确保每个DenseLayer的输入通道数逐渐增加。将新创建的DenseLayer添加为 DenseBlock 的子模块。循环结束后，DenseBlock 就包含了多个DenseLayer，每个DenseLayer都具有逐渐增加的输入通道数量。

class DenseBlock(nn.Sequential):
    def __init__(self, num_layers, num_input_features, bn_size, growth_rate, drop_rate):
        super().__init__()
        for i in range(num_layers):
            layer = DenseLayer(num_input_features + i * growth_rate, growth_rate, bn_size, drop_rate)
            self.add_module("denselayer%d" % (i + 1), layer)

下面是Transition，实现过渡的功能，是在块之间降低通道数量和空间维度。

class Transition(nn.Sequential):
    def __init__(self, num_input_feature, num_output_features):
        super().__init__()
        self.add_module("norm", nn.BatchNorm2d(num_input_feature))
        self.add_module("relu", nn.ReLU(inplace=True))
        self.add_module("conv", nn.Conv2d(num_input_feature, num_output_features, kernel_size=1, stride=1, bias=False))
        self.add_module("pool", nn.AvgPool2d(2, stride=2))

实现深度学习网络主体，将不同的输出特征进行融合，完成分类。

from collections import OrderedDict
class Model(nn.Module):
    def __init__(self, growth_rate=32, block_config=(6, 12, 24, 16), num_init_features=64,
                 bn_size=4, compression_rate=0.5, drop_rate=0, num_classes=1000):
        super().__init__()
        self.features = nn.Sequential(OrderedDict([
            ("conv0", nn.Conv2d(3, num_init_features, kernel_size=7, stride=2, padding=3, bias=False)),
            ("norm0", nn.BatchNorm2d(num_init_features)),
            ("relu0", nn.ReLU(inplace=True)),
            ("pool0", nn.MaxPool2d(3, stride=2, padding=1))
        ]))

        num_features = num_init_features
        for i, num_layers in enumerate(block_config):
            block = DenseBlock(num_layers, num_features, bn_size, growth_rate, drop_rate)
            self.features.add_module("denseblock%d" % (i + 1), block)
            num_features += num_layers * growth_rate
            if i != len(block_config) - 1:
                transition = Transition(num_features, int(num_features * compression_rate))
                self.features.add_module("transition%d" % (i + 1), transition)
                num_features = int(num_features * compression_rate)

        self.features.add_module("norm5", nn.BatchNorm2d(num_features))
        self.features.add_module("relu5", nn.ReLU(inplace=True))

        
        
        self.classifier = nn.Linear(num_features + 2048, num_classes)

        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight)
            elif isinstance(m, nn.BatchNorm2d):
                nn.init.constant_(m.bias, 0)
                nn.init.constant_(m.weight, 1)
            elif isinstance(m, nn.Linear):
                nn.init.constant_(m.bias, 0)

        self.part = ResNetPart()

    def forward(self, x):
        t = self.part(x)
        features = self.features(x)
        out = F.avg_pool2d(features, 7, stride=1).view(features.size(0), -1)
        out = torch.cat([out, t], dim=1) 
        out = self.classifier(out)
        return out

使用summary查看网络

开始训练

定义训练函数

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

    for X, y in dataloader:
        X, y = X.to(device), y.to(device)
        pred = model(X)
        loss = loss_fn(pred, y)

        opt.zero_grad()
        loss.backward()
        opt.step()

        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

定义测试函数

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)
            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

定义学习率、损失函数、优化算法

loss_fn = nn.CrossEntropyLoss()
learn_rate = 0.0002
opt = torch.optim.Adam(model.parameters(), lr=learn_rate)

开始训练，epoch设置为30

import time
epochs = 20
train_loss = []
train_acc = []
test_loss = []
test_acc = []

T1 = time.time()

best_acc = 0
best_model = 0

for epoch in range(epochs):

    model.train()
    epoch_train_acc, epoch_train_loss = train(train_dl, model, loss_fn, opt)
    
    model.eval() # 确保模型不会进行训练操作
    epoch_test_acc, epoch_test_loss = test(test_dl, model, loss_fn)

    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)
    
    print("epoch:%d, train_acc:%.1f%%, train_loss:%.3f, test_acc:%.1f%%, test_loss:%.3f"
          % (epoch + 1, epoch_train_acc * 100, epoch_train_loss, epoch_test_acc * 100, epoch_test_loss))

T2 = time.time()
print('程序运行时间:%s秒' % (T2 - T1))

PATH = './best_model.pth'  # 保存的参数文件名
if best_model is not None:
    torch.save(best_model.state_dict(), PATH)
    print('保存最佳模型')
print("Done")

由于数据量小，所以略微有些过拟合。

可视化

可视化训练过程和测试过程

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()

总结

由于作者水平有限，所以本次模型之间的结合探索采用模型集成，这其中的优势有：1.每个模型通过不同的方式学习数据的表示，将它们集成可以得到更全面、更丰富的特征表示。2.不同的模型可能对数据集中的不同样本和模式有更好的泛化能力，集成可以减少过拟合的风险。这样的模型可以提供更强的鲁棒性，但一定有很多更好的结合方式。