当前位置：首页 > news >正文

深度学习基础案例5--VGG16人脸识别(体验学习的痛苦与乐趣)

news 2024/11/13 11:59:15

🍨 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

前言

这次目标本来要达到60%，但是却非常稳定的达到了40%，😢😢😢😢；
从上个周末到现在，从最初的13%到现在的60%，自己一个人也学习了不少，体验到了期待到无助，又从无助到期待的循环，一个人查阅论文、修改、跑模型验证，反复验证，在这过程，本来在这过程中想好了很多要写的，但是等到真正写的时候，又突然说不出口了，🤠🤠🤠🤠；
最近学校课程多，这周任务又快要结束了，就到后面在不断优化吧，也期待大佬给我提出建议。😢😢😢😢

目标

测试集准确率达到20%

结果

测试集准确率稳定在40%+
没有达到60%+😢😢😢😢😢😢😢😢😢😢😢😢

文章目录

1、模型简介
2、模型训练与优化
- 1、导入数据
- - 1、导入库
  - 2、查看数据类型
  - 3、数据展示
  - 4、数据预处理
  - 5、数据加载与数据划分
- 2、构建VGG-16神经网络模型
- 3、模型训练
- - 1、构建训练函数
  - 2、构建测试集函数
  - 3、设置动态学习率
  - 4、模型正式训练
- 4、结果展示
- 5、预测
- 6、开始优化
- 7、优化一
- 8、优化二
3、总结

1、模型简介

VGG16是一个经典的模型，他在之间广泛用于图像分类的工作，也一直取得了很多人的青睐，它拥有13层卷积，3层池化构成，本文将用VGG16来实现对人脸的识别(本次案例数据1800张)。

VGG16模型结构图如下：

在这里插入图片描述

结合本案例，最总优化的模型结构图如下(论文截图)：

在这里插入图片描述

2、模型训练与优化

1、导入数据

1、导入库

import torch 
import numpy as np 
import torch.nn as nn
import torchvision 
import warnings   # 忽略警告
import os, PIL, pathlib warnings.filterwarnings("ignore")             #忽略警告信息device = ('cuda' if torch.cuda.is_available() else 'cpu')
device

输出：

'cuda'

2、查看数据类型

获取文件夹下的类别名称

data_dir = './data/'
data_dir = pathlib.Path(data_dir)data_path = data_dir.glob('*')
classnames = [str(path).split("\\")[1] for path in data_path]
classnames

输出：

['Angelina Jolie','Brad Pitt','Denzel Washington','Hugh Jackman','Jennifer Lawrence','Johnny Depp','Kate Winslet','Leonardo DiCaprio','Megan Fox','Natalie Portman','Nicole Kidman','Robert Downey Jr','Sandra Bullock','Scarlett Johansson','Tom Cruise','Tom Hanks','Will Smith']

3、数据展示

import matplotlib.pyplot as plt 
from PIL import Image# 文件目录
data_look_dir = './data/AngeLina Jolie/'
# 获得文件名
data_path_list = [f for f in os.listdir(data_look_dir) if f.endswith(('jpg', 'png'))]fig, axes = plt.subplots(2, 8, figsize=(16, 6))  # fig：画板，ases子图# 展示一一部分图片
for ax, img_file in zip(axes.flat, data_path_list):path_name = os.path.join(data_look_dir, img_file)  # 拼接文件目录img = Image.open(path_name)         # 打开文件ax.imshow(img)ax.axis('off')plt.show()

在这里插入图片描述

4、数据预处理

from torchvision import transforms, datasets # 数据预处理，统一格式
data_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] )
])data_all = './data/'total_data = datasets.ImageFolder(root=data_all, transform=data_transform)

5、数据加载与数据划分

# 数据集的划分
train_size = int(len(total_data) * 0.8)
test_size = len(total_data) - train_size
train_data, test_data = torch.utils.data.random_split(total_data, [train_size, test_size])
train_data, test_data

输出：(<torch.utils.data.dataset.Subset at 0x2bd8b922fd0>,<torch.utils.data.dataset.Subset at 0x2bd8b922ac0>)

# 动态加载数据
batch_size = 32 train_dl = torch.utils.data.DataLoader(train_data,batch_size=batch_size,shuffle=True,num_workers=4)test_dl = torch.utils.data.DataLoader(test_data,batch_size=batch_size,shuffle=True,num_workers=4)

# 查看处理后图片格式
for format, data in test_dl:print("format: ", format.shape)print("data: ", data)break

format:  torch.Size([32, 3, 224, 224])
data:  tensor([14, 12,  4, 13,  6,  3, 13,  4, 16,  0,  2,  4,  5,  6,  7,  2,  6,  0,14, 10, 13,  3,  8,  7, 10,  4, 12,  0,  4, 14,  3, 15])

2、构建VGG-16神经网络模型

from torchvision.models import vgg16# 加载模型
model = vgg16(pretrained=False).to(device)# vgg16已经通过大量的模型训练，故不需要参数更新
for param in model.parameters():param.required_grad = False     # 禁止梯度更新# 修改全连接层
model.classifier._modules['6'] = nn.Linear(4096, len(classnames))
model.to(device)
model

输出：

VGG((features): Sequential((0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(1): ReLU(inplace=True)(2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(3): ReLU(inplace=True)(4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)(5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(6): ReLU(inplace=True)(7): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(8): ReLU(inplace=True)(9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)(10): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(11): ReLU(inplace=True)(12): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(13): ReLU(inplace=True)(14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(15): ReLU(inplace=True)(16): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)(17): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(18): ReLU(inplace=True)(19): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(20): ReLU(inplace=True)(21): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(22): ReLU(inplace=True)(23): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)(24): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(25): ReLU(inplace=True)(26): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(27): ReLU(inplace=True)(28): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(29): ReLU(inplace=True)(30): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False))(avgpool): AdaptiveAvgPool2d(output_size=(7, 7))(classifier): Sequential((0): Linear(in_features=25088, out_features=4096, bias=True)(1): ReLU(inplace=True)(2): Dropout(p=0.5, inplace=False)(3): Linear(in_features=4096, out_features=4096, bias=True)(4): ReLU(inplace=True)(5): Dropout(p=0.5, inplace=False)(6): Linear(in_features=4096, out_features=17, bias=True))
)

3、模型训练

1、构建训练函数

def train(dataloader, model, loss_fn, optimizer):size = len(dataloader.dataset)batch_size = len(dataloader)train_acc, train_loss = 0, 0# 模型预测for X, y in dataloader:X, y = X.to(device), y.to(device)# 预测pred = model(X)loss = loss_fn(pred, y)# 梯度更新optimizer.zero_grad()loss.backward()optimizer.step()# 计算损失和准确率train_loss += losstrain_acc += (pred.argmax(1) == y).type(torch.float64).sum().item()# 计算总和train_acc /= sizetrain_loss /= batch_sizereturn train_acc, train_loss

2、构建测试集函数

def test(dataloader, model, loss_fn):size = len(dataloader.dataset)batch_size = 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_loss += loss.item()test_acc += (pred.argmax(1) == y).type(torch.float64).sum().item()test_acc /= size test_loss /= batch_sizereturn test_acc, test_loss

3、设置动态学习率

learn_rate = 1e-4 
loss_fn = nn.CrossEntropyLoss()
func = lambda epoch : (0.92 ** (epoch // 2))
optimizer = torch.optim.SGD(model.parameters(), lr=learn_rate)
scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=func)

4、模型正式训练

import copy train_acc = []
train_loss = []
test_acc = []
test_loss = []epoches = 40best_acc = 0   # 最佳学习率for epoch in range(epoches):model.train()epoch_train_acc, epoch_trian_loss = train(train_dl, model, loss_fn, optimizer)# 动态更新学习率scheduler.step()model.eval()epoch_test_acc, epoch_test_loss = test(test_dl, model, loss_fn)# 保存最佳参数模型if epoch_test_acc > best_acc:best_acc = epoch_test_accbest_model = copy.deepcopy(model)train_acc.append(epoch_train_acc)train_loss.append(epoch_trian_loss)test_acc.append(epoch_test_acc)test_loss.append(epoch_test_loss)# 保存当前学习率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_trian_loss,  epoch_test_acc*100, epoch_test_loss, lr))# 保存最佳模型到文件
path = './best_model.pth'
torch.save(model.state_dict(), path)

输出：
Epoch: 1, Train_acc:5.3%, Train_loss:2.835, Test_acc:5.0%, Test_loss:2.838, Lr:1.00E-04
Epoch: 2, Train_acc:6.0%, Train_loss:2.838, Test_acc:4.4%, Test_loss:2.835, Lr:9.20E-05
Epoch: 3, Train_acc:6.0%, Train_loss:2.837, Test_acc:4.4%, Test_loss:2.835, Lr:9.20E-05
Epoch: 4, Train_acc:5.5%, Train_loss:2.837, Test_acc:4.4%, Test_loss:2.835, Lr:8.46E-05
Epoch: 5, Train_acc:5.1%, Train_loss:2.837, Test_acc:4.4%, Test_loss:2.834, Lr:8.46E-05
Epoch: 6, Train_acc:6.6%, Train_loss:2.833, Test_acc:3.9%, Test_loss:2.834, Lr:7.79E-05
Epoch: 7, Train_acc:6.2%, Train_loss:2.836, Test_acc:3.9%, Test_loss:2.831, Lr:7.79E-05
Epoch: 8, Train_acc:6.8%, Train_loss:2.829, Test_acc:3.9%, Test_loss:2.834, Lr:7.16E-05
Epoch: 9, Train_acc:6.4%, Train_loss:2.834, Test_acc:3.9%, Test_loss:2.832, Lr:7.16E-05
Epoch:10, Train_acc:6.7%, Train_loss:2.833, Test_acc:3.9%, Test_loss:2.832, Lr:6.59E-05
Epoch:11, Train_acc:5.6%, Train_loss:2.833, Test_acc:3.9%, Test_loss:2.832, Lr:6.59E-05
Epoch:12, Train_acc:5.8%, Train_loss:2.835, Test_acc:3.9%, Test_loss:2.832, Lr:6.06E-05
Epoch:13, Train_acc:6.2%, Train_loss:2.832, Test_acc:4.2%, Test_loss:2.830, Lr:6.06E-05
Epoch:14, Train_acc:6.7%, Train_loss:2.832, Test_acc:4.2%, Test_loss:2.830, Lr:5.58E-05
Epoch:15, Train_acc:6.2%, Train_loss:2.834, Test_acc:4.2%, Test_loss:2.831, Lr:5.58E-05
Epoch:16, Train_acc:7.4%, Train_loss:2.828, Test_acc:5.0%, Test_loss:2.828, Lr:5.13E-05
Epoch:17, Train_acc:6.9%, Train_loss:2.834, Test_acc:6.1%, Test_loss:2.831, Lr:5.13E-05
Epoch:18, Train_acc:7.1%, Train_loss:2.834, Test_acc:8.1%, Test_loss:2.829, Lr:4.72E-05
Epoch:19, Train_acc:6.7%, Train_loss:2.833, Test_acc:8.9%, Test_loss:2.829, Lr:4.72E-05
Epoch:20, Train_acc:7.5%, Train_loss:2.829, Test_acc:9.2%, Test_loss:2.827, Lr:4.34E-05
Epoch:21, Train_acc:8.3%, Train_loss:2.831, Test_acc:11.1%, Test_loss:2.828, Lr:4.34E-05
Epoch:22, Train_acc:6.2%, Train_loss:2.830, Test_acc:11.7%, Test_loss:2.828, Lr:4.00E-05
Epoch:23, Train_acc:8.5%, Train_loss:2.828, Test_acc:12.5%, Test_loss:2.829, Lr:4.00E-05
Epoch:24, Train_acc:7.3%, Train_loss:2.830, Test_acc:11.9%, Test_loss:2.827, Lr:3.68E-05
Epoch:25, Train_acc:8.4%, Train_loss:2.827, Test_acc:11.9%, Test_loss:2.828, Lr:3.68E-05
Epoch:26, Train_acc:8.5%, Train_loss:2.830, Test_acc:12.8%, Test_loss:2.828, Lr:3.38E-05
Epoch:27, Train_acc:6.9%, Train_loss:2.831, Test_acc:13.1%, Test_loss:2.828, Lr:3.38E-05
Epoch:28, Train_acc:7.8%, Train_loss:2.828, Test_acc:13.1%, Test_loss:2.828, Lr:3.11E-05
Epoch:29, Train_acc:7.8%, Train_loss:2.826, Test_acc:13.1%, Test_loss:2.828, Lr:3.11E-05
Epoch:30, Train_acc:6.6%, Train_loss:2.831, Test_acc:13.1%, Test_loss:2.827, Lr:2.86E-05
Epoch:31, Train_acc:8.4%, Train_loss:2.828, Test_acc:13.1%, Test_loss:2.828, Lr:2.86E-05
Epoch:32, Train_acc:9.4%, Train_loss:2.828, Test_acc:13.1%, Test_loss:2.829, Lr:2.63E-05
Epoch:33, Train_acc:9.1%, Train_loss:2.829, Test_acc:13.1%, Test_loss:2.826, Lr:2.63E-05
Epoch:34, Train_acc:8.2%, Train_loss:2.828, Test_acc:13.1%, Test_loss:2.826, Lr:2.42E-05
Epoch:35, Train_acc:9.8%, Train_loss:2.829, Test_acc:13.1%, Test_loss:2.827, Lr:2.42E-05
Epoch:36, Train_acc:7.9%, Train_loss:2.827, Test_acc:13.1%, Test_loss:2.826, Lr:2.23E-05
Epoch:37, Train_acc:9.4%, Train_loss:2.827, Test_acc:13.1%, Test_loss:2.825, Lr:2.23E-05
Epoch:38, Train_acc:8.2%, Train_loss:2.829, Test_acc:13.1%, Test_loss:2.826, Lr:2.05E-05
Epoch:39, Train_acc:8.8%, Train_loss:2.827, Test_acc:13.1%, Test_loss:2.827, Lr:2.05E-05
Epoch:40, Train_acc:7.6%, Train_loss:2.826, Test_acc:13.1%, Test_loss:2.824, Lr:1.89E-05

不知道为什么，有时候这里train_loss还存储在GPU中，故需要转换：

if isinstance(train_loss, torch.Tensor) and train_loss.device.type == 'cuda':train_loss = train_loss.cpu()
# 转换
train_loss_cpu = [t.detach().cpu().numpy() for t in train_loss if isinstance(t, torch.Tensor)]# 重新开始
train_loss = np.array(train_loss_cpu)

4、结果展示

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plt.figure(figsize=(12, 3))epoch_range = range(epoches)plt.subplot(1, 2, 1)
plt.plot(epoch_range, train_acc, label='Train Accurary')
plt.plot(epoch_range, test_acc, label='Test Accurary')
plt.title('Accurary')
plt.legend(loc='lower right')plt.subplot(1, 2, 2)
plt.plot(epoch_range, train_loss, label='Train Loss')
plt.plot(epoch_range, test_loss, label='Test Loss')
plt.title('Loss')
plt.legend(loc='upper right')plt.show()

在这里插入图片描述

5、预测

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')  # RGB格式打开plt.imshow(test_img)  # 展示预测图片# 压缩图片，更好训练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('预测结果：', pred_class)

predict_one_image('./data/Angelina Jolie/001_fe3347c0.jpg', model, data_transform, classes)

预测结果： Scarlett Johansson

在这里插入图片描述

6、开始优化

本次实验总结：

准确率：
- 训练集：刚开始极低，后面更是一直不变
- 测试集：虽然有逐步上升趋势，但是却一直很低
损失率：
- 测试集和训练集的损失率变化大差不差，都是逐渐减低，但是降低极少，不到0.1，效果不好
最后：总的来说，直接调用VGG的模型，跑出来的效果极差

7、优化一

通过查阅论文、相关博客，最后对以上模型进行了以下优化：

全连接层：减少层数，降低神经元个数，添加Dropout层，修改代码如下：

model.classifier = nn.Sequential(nn.Linear(512 * 7 * 7, 1024),nn.ReLU(inplace=True),nn.Dropout(0.5,inplace=False),nn.Linear(1024,512),nn.ReLU(inplace=True),nn.Dropout(0.5),nn.Linear(512,len(classnames))) # 修改vgg16模型中最后一层全连接层，输出目标类别个数

学习率：由于在模型训练中经常出现准确率不变的情况，故降低学习率，1e-4 --> 1e-3，同时变小自动调整学习率速度，0.88 ** (epoch // 2)
优化器：通过查阅论发现优化器不同也会导致不同结果，实际也得到了验证，SGD—>Adam
图像增强：由于数据量少，故对图像进行随机旋转，以增强数据

最后，通过训练，结果如图：

在这里插入图片描述

效果比上一个模型好很多，训练集准确率来到了40%，但是在后面出现了一个现象，就是训练集准确率一直升高，达到了90%以上，但是测试集准确率却一直没有怎么变化，所以，就出现了到后面测试集的损失值上升的趋势，模型不稳定。

8、优化二

后面继续查阅论文，最后索性，将全部模型进行修改，添加BN层、将激活函数变成LeakyReLU，同时减少全连接层的数量，最后模型效果得到了稳定的提升，但是训练集和测试集的准确率上升一直变得很缓慢，上升不去，但是相比于优化一来说，效果好了很多，结果图如下：

在这里插入图片描述

优化代码：

class Work_Net(nn.Module):def __init__(self):super(Work_Net, self).__init__()# Block1self.block1 = nn.Sequential(nn.Conv2d(3, 64, kernel_size=3, padding=1),nn.BatchNorm2d(64),   # 1、添加BN层nn.LeakyReLU(negative_slope=0.01),  # 2、修改激活函数nn.Conv2d(64, 64, kernel_size=3, padding=1),nn.BatchNorm2d(64),   # 1、添加BN层nn.LeakyReLU(negative_slope=0.1),  # 2、修改激活函数nn.MaxPool2d(kernel_size=2, stride=2))# Block 2self.block2 = nn.Sequential(nn.Conv2d(64, 128, kernel_size=3, padding=1),nn.BatchNorm2d(128),   # 1、添加BN层nn.LeakyReLU(negative_slope=0.01),  # 2、修改激活函数nn.Conv2d(128, 128, kernel_size=3, padding=1),nn.BatchNorm2d(128),   # 1、添加BN层nn.LeakyReLU(negative_slope=0.01),  # 2、修改激活函数nn.MaxPool2d(kernel_size=2, stride=2)    )# Block3self.block3 = nn.Sequential(nn.Conv2d(128, 256, kernel_size=3, padding=1),nn.BatchNorm2d(256),   # 1、添加BN层nn.LeakyReLU(negative_slope=0.01),  # 2、修改激活函数nn.Conv2d(256, 256, kernel_size=3, padding=1),nn.BatchNorm2d(256),   # 1、添加BN层nn.LeakyReLU(negative_slope=0.01),  # 2、修改激活函数nn.Conv2d(256, 256, kernel_size=3, padding=1),nn.BatchNorm2d(256),   # 1、添加BN层nn.LeakyReLU(negative_slope=0.01),  # 2、修改激活函数nn.MaxPool2d(kernel_size=2, stride=2))# Block 4self.block4 = nn.Sequential(nn.Conv2d(256, 512, kernel_size=3, padding=1),nn.BatchNorm2d(512),   # 1、添加BN层nn.LeakyReLU(negative_slope=0.01),  # 2、修改激活函数nn.Conv2d(512, 512, kernel_size=3, padding=1),nn.BatchNorm2d(512),   # 1、添加BN层nn.LeakyReLU(negative_slope=0.01),  # 2、修改激活函数nn.Conv2d(512, 512, kernel_size=3, padding=1),nn.BatchNorm2d(512),   # 1、添加BN层nn.LeakyReLU(negative_slope=0.01),  # 2、修改激活函数nn.MaxPool2d(kernel_size=2, stride=2))# Block5self.block5 = nn.Sequential(nn.Conv2d(512, 512, kernel_size=3, padding=1),nn.BatchNorm2d(512),   # 1、添加BN层nn.LeakyReLU(negative_slope=0.01),  # 2、修改激活函数nn.Conv2d(512, 512, kernel_size=3, padding=1),nn.BatchNorm2d(512),   # 1、添加BN层nn.LeakyReLU(negative_slope=0.01),  # 2、修改激活函数nn.Conv2d(512, 512, kernel_size=3, padding=1),nn.BatchNorm2d(512),   # 1、添加BN层nn.LeakyReLU(negative_slope=0.01),  # 2、修改激活函数nn.MaxPool2d(kernel_size=2, stride=2))# Block6self.classifier = nn.Sequential(nn.Linear(512 * 7 * 7, 1024),nn.LeakyReLU(negative_slope=0.01),  # 2、修改激活函数nn.Dropout(0.5),nn.Linear(1024, 512),nn.LeakyReLU(negative_slope=0.01),  # 2、修改激活函数nn.Dropout(0.5),nn.Linear(512, len(classnames)))def forward(self, x):x = self.block1(x)x = self.block2(x)x = self.block3(x)x = self.block4(x)x = self.block5(x)x = x.view(x.size(0), -1)x = self.classifier(x)return x

3、总结

这一次虽然没有达到目标(测试集准确率60%)，但是对深度学习的概念有了更加清晰的认识：

全连接层：全连接层是通过CNN训练得到的结果后，将训练结果进行展开，然后根据目标类型进行分类，如果全连接层展开很大的话，降低神经元数量，增加Dropout层，可以防止过拟合，防止出现测试集、训练集的准确率或者损失率一直不变的情况；
激活函数：常用的激活函数主要是ReLU，但是在数据量少的时候也可以尝试用LeaykReLU，考虑特征值计算的时候出现复数的情况；
优化器：不同优化器可能导致不同的效果，需要结合数据运算的效果来选取；
学习率：过大的学习率会收敛过快，提取不到有效的信息，但是学习率过小，极容易出现训练集、测试集的准确率或者学习率不变的情况；
数据增强：数据量少的时候可以通过数据增强，可以通过transforms.Compose以增加数据

查看全文

http://www.mrgr.cn/news/30478.html