致谢&＃xff1a;霹雳吧啦Wz&＃xff1a;https://space.bilibili.com/18161609

目录

致谢&＃xff1a;霹雳吧啦Wz&＃xff1a;https://space.bilibili.com/18161609

1 本次要点

1.1 深度学习理论

1.2 pytorch框架语法

2 网络简介

2.1 历史意义

2.2 网络亮点

2.3 网络架构

3 代码结构

3.1 model.py

3.2 train.py

3.3 predict.py

3.4 split_data.py

1 本次要点

1.1 深度学习理论

1. 经过一次卷积操作后&＃xff0c;图像新尺寸计算公式&＃xff1a;&＃xff08;如果padding [p1, p2]中p1,p2不相等&＃xff0c;那么公式中2P就变为P1&＃43;P2&＃xff09;&＃xff08;如果结果值不是整数&＃xff0c;pytorch中会自动忽略最后一行以及最后一列&＃xff0c;以保证N为整数。&＃xff09;
2. 
3.  

1.2 pytorch框架语法

1. pytorch可以自定义网络权重的初始化方法&＃xff08;见model.py&＃xff09;。

2. pata &＃61; list(net.parameters()) #查看模型参数

2 网络简介

2.1 历史意义

• 2012年ImageNet图像分类冠军网络&＃xff0c;分类准确率由传统的 70%&＃43;直接提升到 80%&＃43;。在那年之后&＃xff0c;深
  度学习开始迅速发展。

2.2 网络亮点

1. 首次利用 GPU 进行网络加速训练。
2. 使用了 ReLU 激活函数&＃xff0c;而不是传统的 Sigmoid 激活函数以及 Tanh 激活函数。
3. 在前两层的全连接层中使用了 Dropout 随机失活神经元操作&＃xff0c;以减少过拟合。

2.3 网络架构

备注&＃xff1a;padding: [1, 2]即图像最左边缘加1列0&＃xff0c;最右边缘加2列0。图像最上边缘加1行0&＃xff0c;图像最下边缘加2行0。

3 代码结构

• model.py
• train.py
• predict.py
• split_data.py&＃xff08;数据集划分&＃xff09;

3.1 model.py

import torch.nn as nn
import torch"""
本AlexNet复现相比原论文&＃xff0c;每层的卷积核个数减半。
"""
class AlexNet(nn.Module):def __init__(self, num_classes&＃61;1000, init_weights&＃61;False):super(AlexNet, self).__init__()# nn.Sequential():将一系列层结构进行打包。省去每一层都用一个变量去表示。self.features &＃61; nn.Sequential(nn.Conv2d(3, 48, kernel_size&＃61;11, stride&＃61;4, padding&＃61;2), # input[3, 224, 224] output[48, 55, 55]nn.ReLU(inplace&＃61;True), #inplace&＃xff1a;通过增加计算量来降低内存使用&＃xff0c;从而可以载入更大模型&＃xff08;默认False&＃xff09;。nn.MaxPool2d(kernel_size&＃61;3, stride&＃61;2), # output[48, 27, 27]nn.Conv2d(48, 128, kernel_size&＃61;5, padding&＃61;2), # output[128, 27, 27]nn.ReLU(inplace&＃61;True),nn.MaxPool2d(kernel_size&＃61;3, stride&＃61;2), # output[128, 13, 13]nn.Conv2d(128, 192, kernel_size&＃61;3, padding&＃61;1), # output[192, 13, 13]nn.ReLU(inplace&＃61;True),nn.Conv2d(192, 192, kernel_size&＃61;3, padding&＃61;1), # output[192, 13, 13]nn.ReLU(inplace&＃61;True),nn.Conv2d(192, 128, kernel_size&＃61;3, padding&＃61;1), # output[128, 13, 13]nn.ReLU(inplace&＃61;True),nn.MaxPool2d(kernel_size&＃61;3, stride&＃61;2), # output[128, 6, 6])self.classifier &＃61; nn.Sequential(nn.Dropout(p&＃61;0.5),nn.Linear(128 * 6 * 6, 2048), # 输入&＃xff1a;128通道*6*6&＃xff08;特征图大小&＃xff09;&＃xff08;到此之前会拉成1维&＃xff09;nn.ReLU(inplace&＃61;True),nn.Dropout(p&＃61;0.5),nn.Linear(2048, 2048),nn.ReLU(inplace&＃61;True),nn.Linear(2048, num_classes),)if init_weights:self._initialize_weights()def forward(self, x):x &＃61; self.features(x)x &＃61; torch.flatten(x, start_dim&＃61;1) # torch中顺序[B,C,H,W]&＃xff0c;start_dim&＃61;1就是将C维度拉平。x &＃61; self.classifier(x)return x# 初始化权重方式&＃xff08;框架有默认&＃xff0c;如果要自定义可如下方式写&＃xff09;def _initialize_weights(self):for m in self.modules():if isinstance(m, nn.Conv2d):nn.init.kaiming_normal_(m.weight, mode&＃61;&＃39;fan_out&＃39;, nonlinearity&＃61;&＃39;relu&＃39;)if m.bias is not None:nn.init.constant_(m.bias, 0)elif isinstance(m, nn.Linear):nn.init.normal_(m.weight, 0, 0.01)nn.init.constant_(m.bias, 0)

3.2 train.py

import torch
import torch.nn as nn
from torchvision import transforms, datasets, utils
import matplotlib.pyplot as plt
import numpy as np
import torch.optim as optim
from model import AlexNet
import os
import json
import time"""
数据集&＃xff1a;花分类&＃xff08;5类&＃xff09;
"""def main():device &＃61; torch.device("cuda:0" if torch.cuda.is_available() else "cpu")print("using {} device.".format(device))data_transform &＃61; {"train": transforms.Compose([transforms.RandomResizedCrop(224),transforms.RandomHorizontalFlip(),#水平随机翻转transforms.ToTensor(),transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]),"val": transforms.Compose([transforms.Resize((224, 224)), # cannot 224, must (224, 224)transforms.ToTensor(),transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])}data_root &＃61; os.path.abspath(os.path.join(os.getcwd(), "../..")) #os.getcwd()&＃xff1a;获取当前绝对路径。"../.."返回到上上层路径。image_path &＃61; os.path.join(data_root, "data_set", "flower_data") # flower data set pathassert os.path.exists(image_path), "{} path does not exist.".format(image_path)train_dataset &＃61; datasets.ImageFolder(root&＃61;os.path.join(image_path, "train"),transform&＃61;data_transform["train"])train_num &＃61; len(train_dataset)# {&＃39;daisy&＃39;:0, &＃39;dandelion&＃39;:1, &＃39;roses&＃39;:2, &＃39;sunflower&＃39;:3, &＃39;tulips&＃39;:4}flower_list &＃61; train_dataset.class_to_idxcla_dict &＃61; dict((val, key) for key, val in flower_list.items())#将键和值顺序反过来。目的是让模型预测的结果索引&＃xff0c;可直接找到对应的类型。# write dict into json filejson_str &＃61; json.dumps(cla_dict, indent&＃61;4)#编码成json格式with open(&＃39;class_indices.json&＃39;, &＃39;w&＃39;) as json_file:#新建json文件并写入内容json_file.write(json_str)batch_size &＃61; 32nw &＃61; min([os.cpu_count(), batch_size if batch_size > 1 else 0, 8]) # number of workersprint(&＃39;Using {} dataloader workers every process&＃39;.format(nw))train_loader &＃61; torch.utils.data.DataLoader(train_dataset,batch_size&＃61;batch_size, shuffle&＃61;True,num_workers&＃61;nw)validate_dataset &＃61; datasets.ImageFolder(root&＃61;os.path.join(image_path, "val"),transform&＃61;data_transform["val"])val_num &＃61; len(validate_dataset)validate_loader &＃61; torch.utils.data.DataLoader(validate_dataset,batch_size&＃61;4, shuffle&＃61;False,num_workers&＃61;nw)print("using {} images for training, {} images fot validation.".format(train_num,# 查看数据集代码 val_num))# test_data_iter &＃61; iter(validate_loader)# test_image, test_label &＃61; test_data_iter.next()## def imshow(img):# img &＃61; img / 2 &＃43; 0.5 # unnormalize# npimg &＃61; img.numpy()# plt.imshow(np.transpose(npimg, (1, 2, 0)))# plt.show()## print(&＃39; &＃39;.join(&＃39;%5s&＃39; % cla_dict[test_label[j].item()] for j in range(4)))# imshow(utils.make_grid(test_image))net &＃61; AlexNet(num_classes&＃61;5, init_weights&＃61;True)net.to(device)loss_function &＃61; nn.CrossEntropyLoss()# pata &＃61; list(net.parameters()) #查看模型参数&＃xff08;调试用&＃xff09;optimizer &＃61; optim.Adam(net.parameters(), lr&＃61;0.0002)save_path &＃61; &＃39;./AlexNet.pth&＃39;best_acc &＃61; 0.0for epoch in range(10):# 训练阶段net.train() #自动判定dropout或BN层是否应该启用。running_loss &＃61; 0.0t1 &＃61; time.perf_counter()for step, data in enumerate(train_loader, start&＃61;0):images, labels &＃61; dataoptimizer.zero_grad()outputs &＃61; net(images.to(device))loss &＃61; loss_function(outputs, labels.to(device))loss.backward()#反向传播optimizer.step()#更新每个节点参数# print statisticsrunning_loss &＃43;&＃61; loss.item()# print train process 打印训练信息rate &＃61; (step &＃43; 1) / len(train_loader)a &＃61; "*" * int(rate * 50)b &＃61; "." * int((1 - rate) * 50)print("\rtrain loss: {:^3.0f}%[{}->{}]{:.3f}".format(int(rate * 100), a, b, loss), end&＃61;"")print()print(time.perf_counter()-t1)# 验证阶段net.eval() #自动判定dropout或BN层是否应该启用。acc &＃61; 0.0 # accumulate accurate number / epochwith torch.no_grad():#不去计算损失梯度for val_data in validate_loader:val_images, val_labels &＃61; val_dataoutputs &＃61; net(val_images.to(device))predict_y &＃61; torch.max(outputs, dim&＃61;1)[1]acc &＃43;&＃61; (predict_y &＃61;&＃61; val_labels.to(device)).sum().item()val_accurate &＃61; acc / val_numif val_accurate > best_acc: best_acc &＃61; val_accuratetorch.save(net.state_dict(), save_path)print(&＃39;[epoch %d] train_loss: %.3f test_accuracy: %.3f&＃39; %(epoch &＃43; 1, running_loss / step, val_accurate))print(&＃39;Finished Training&＃39;)if __name__ &＃61;&＃61; &＃39;__main__&＃39;:main()

训练结果&＃xff1a;

3.3 predict.py

import torch
from model import AlexNet
from PIL import Image
from torchvision import transforms
import matplotlib.pyplot as plt
import jsondata_transform &＃61; transforms.Compose([transforms.Resize((224, 224)),transforms.ToTensor(),transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])# load image
img &＃61; Image.open("../tulip.jpg")
plt.imshow(img)
# [N, C, H, W]
img &＃61; data_transform(img)
# expand batch dimension
img &＃61; torch.unsqueeze(img, dim&＃61;0)# read class_indict
try:json_file &＃61; open(&＃39;./class_indices.json&＃39;, &＃39;r&＃39;)class_indict &＃61; json.load(json_file)
except Exception as e:print(e)exit(-1)# create model
model &＃61; AlexNet(num_classes&＃61;5)
# load model weights
model_weight_path &＃61; "./AlexNet.pth"
model.load_state_dict(torch.load(model_weight_path))
model.eval()
with torch.no_grad():#不去计算损失梯度# predict classoutput &＃61; torch.squeeze(model(img))#torch.squeeze()&＃xff1a;对数据的维度进行压缩&＃xff0c;去掉维数为1的的维度predict &＃61; torch.softmax(output, dim&＃61;0)#将预测结果值转换为概率分布形式。predict_cla &＃61; torch.argmax(predict).numpy()
print(class_indict[str(predict_cla)], predict[predict_cla].item())
plt.show()

输出&＃xff1a;

3.4 split_data.py

import os
from shutil import copy, rmtree
import random"""
使用步骤如下&＃xff1a;
&＃xff08;1&＃xff09;在data_set文件夹下创建新文件夹"flower_data"
&＃xff08;2&＃xff09;点击链接下载花分类数据集 http://download.tensorflow.org/example_images/flower_photos.tgz
&＃xff08;3&＃xff09;解压数据集到flower_data文件夹下
&＃xff08;4&＃xff09;执行"split_data.py"脚本自动将数据集划分成训练集train和验证集val├── flower_data ├── flower_photos&＃xff08;解压的数据集文件夹&＃xff0c;3670个样本&＃xff09; ├── train&＃xff08;生成的训练集&＃xff0c;3306个样本&＃xff09; └── val&＃xff08;生成的验证集&＃xff0c;364个样本&＃xff09;
"""def mk_file(file_path: str):if os.path.exists(file_path):# 如果文件夹存在&＃xff0c;则先删除原文件夹在重新创建rmtree(file_path)os.makedirs(file_path)def main():# 保证随机可复现random.seed(0)# 将数据集中10%的数据划分到验证集中split_rate &＃61; 0.1# 指向你解压后的flower_photos文件夹cwd &＃61; os.getcwd()data_root &＃61; os.path.join(cwd, "flower_data")origin_flower_path &＃61; os.path.join(data_root, "flower_photos")assert os.path.exists(origin_flower_path)flower_class &＃61; [cla for cla in os.listdir(origin_flower_path)if os.path.isdir(os.path.join(origin_flower_path, cla))]# 建立保存训练集的文件夹train_root &＃61; os.path.join(data_root, "train")mk_file(train_root)for cla in flower_class:# 建立每个类别对应的文件夹mk_file(os.path.join(train_root, cla))# 建立保存验证集的文件夹val_root &＃61; os.path.join(data_root, "val")mk_file(val_root)for cla in flower_class:# 建立每个类别对应的文件夹mk_file(os.path.join(val_root, cla))for cla in flower_class:cla_path &＃61; os.path.join(origin_flower_path, cla)images &＃61; os.listdir(cla_path)num &＃61; len(images)# 随机采样验证集的索引eval_index &＃61; random.sample(images, k&＃61;int(num*split_rate))for index, image in enumerate(images):if image in eval_index:# 将分配至验证集中的文件复制到相应目录image_path &＃61; os.path.join(cla_path, image)new_path &＃61; os.path.join(val_root, cla)copy(image_path, new_path)else:# 将分配至训练集中的文件复制到相应目录image_path &＃61; os.path.join(cla_path, image)new_path &＃61; os.path.join(train_root, cla)copy(image_path, new_path)print("\r[{}] processing [{}/{}]".format(cla, index&＃43;1, num), end&＃61;"") # processing barprint()print("processing done!")if __name__ &＃61;&＃61; &＃39;__main__&＃39;:main()

输出&＃xff1a;