深度学习P5：运动鞋识别

最新推荐文章于 2024-07-25 23:06:58 发布

XiaoLaJi2019

最新推荐文章于 2024-07-25 23:06:58 发布

阅读量853

点赞数 31

分类专栏：慢慢深度学习路文章标签：深度学习 pytorch cnn

本文链接：https://blog.csdn.net/XiaoLaJi2019/article/details/136951181

版权

慢慢深度学习路专栏收录该内容

1 篇文章 0 订阅

订阅专栏

>- **🍨 本文为[🔗365天深度学习训练营](https://mp.weixin.qq.com/s/rbOOmire8OocQ90QM78DRA) 中的学习记录博客**
>- **🍖 原作者：[K同学啊 | 接辅导、项目定制](https://mtyjkh.blog.csdn.net/)**

前期准备

导入库

import torch
import torch.nn as nn
import torchvision.transforms as transforms
import torchvision
from torchvision import transforms, datasets

import os,PIL,pathlib

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

device

导入数据

import os,PIL,random,pathlib

data_dir = './5-data/'
data_dir = pathlib.Path(data_dir)

data_paths  = list(data_dir.glob('*'))
classeNames = [str(path).split("\\")[1] for path in data_paths]
classeNames

值得一提的是此处用“\”分割后获取的是5-data这个目录下一级目录的名称，获取目录名称可以根据实际调整

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转换为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], 
        std=[0.229, 0.224, 0.225])  # 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。
])

test_transform = transforms.Compose([
    transforms.Resize([224, 224]),  # 将输入图片resize成统一尺寸
    transforms.ToTensor(),          # 将PIL Image或numpy.ndarray转换为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], 
        std=[0.229, 0.224, 0.225])  # 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。
])

train_dataset = datasets.ImageFolder("./5-data/train/",transform=train_transforms)
test_dataset  = datasets.ImageFolder("./5-data/test/",transform=train_transforms)

用Pytorch分析图片的常规预处理，注释掉的翻转其实是训练数据前的常见操作，一些翻转、裁剪等可能更有利于模型学习图片特征，根据具体情况可尝试是否能提升模型的准确率。此数据集为深深度学习经典数据集，因此mean和std直接套用，位置数据标准化需要自行抽样计算。

train_dataset.class_to_idx

{'adidas': 0, 'nike': 1}

Pytorch组件，图片分类任务时可将类别名称映射为唯一的整数索引，类型是字典

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)

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

常规的批量加载数据，打乱，并行设置，最后打印shape便于后续架构模型

模型构建

构建CNN网络

import torch.nn.functional as F

class Model(nn.Module):
    def __init__(self):
        super(Model, self).__init__()
        self.conv1=nn.Sequential(
            nn.Conv2d(3, 12, kernel_size=5, padding=0), # 12*220*220
            nn.BatchNorm2d(12),
            nn.ReLU())
        
        self.conv2=nn.Sequential(
            nn.Conv2d(12, 12, kernel_size=5, padding=0), # 12*216*216
            nn.BatchNorm2d(12),
            nn.ReLU())
        
        self.pool3=nn.Sequential(
            nn.MaxPool2d(2))                              # 12*108*108
        
        self.conv4=nn.Sequential(
            nn.Conv2d(12, 24, kernel_size=5, padding=0), # 24*104*104
            nn.BatchNorm2d(24),
            nn.ReLU())
        
        self.conv5=nn.Sequential(
            nn.Conv2d(24, 24, kernel_size=5, padding=0), # 24*100*100
            nn.BatchNorm2d(24),
            nn.ReLU())
        
        self.pool6=nn.Sequential(
            nn.MaxPool2d(2))                              # 24*50*50

        self.dropout = nn.Sequential(
            nn.Dropout(0.2))
        
        self.fc=nn.Sequential(
            nn.Linear(24*50*50, len(classeNames)))
        
    def forward(self, x):
        
        batch_size = x.size(0)
        x = self.conv1(x)  # 卷积-BN-激活
        x = self.conv2(x)  # 卷积-BN-激活
        x = self.pool3(x)  # 池化
        x = self.conv4(x)  # 卷积-BN-激活
        x = self.conv5(x)  # 卷积-BN-激活
        x = self.pool6(x)  # 池化
        x = self.dropout(x)
        x = x.view(batch_size, -1)  # flatten 变成全连接网络需要的输入 (batch, 24*50*50) ==> (batch, -1), -1 此处自动算出的是24*50*50
        x = self.fc(x)
       
        return x

device = "cuda" if torch.cuda.is_available() else "cpu"
print("Using {} device".format(device))

model = Model().to(device)
model

两个需要注意的地方

1、ReLU激活函数f(x)=max(0,x)。在前面的CNN网络中并没有设计该激活函数，该函数的主要作用是通过引入非线性，提高了神经网络模型的学习能力和表达能力。

（1）增加模型表达力和学习力。若没有该函数或激活函数只是线性的，模型无论多少层输出的都是输入的线性组合。非线性激活函数使得神经网络能够学习复杂的函数映射。理论上，一个带有至少一层隐藏层和足够数量非线性激活函数的神经网络可以近似任何连续函数，这被称为神经网络的万能逼近定理。

（2）解决梯度消失问题。一些非线性激活函数（如ReLU）在正区间内的导数不为零（对ReLU而言，正区间的导数为1）。这有助于减少在深层网络中反向传播时梯度消失的问题，因为梯度不会随着网络深度的增加而急剧减小。

（3）增强特征表达。引入稀疏性，ReLU激活函数会将所有负值输出为0，这导致网络在任何时候只有一部分神经元被激活。这种稀疏激活可以提高网络的计算效率，并且有助于减少过拟合。同时，稀疏激活意味着网络能够在较少的特征上集中学习，这可以帮助模型捕捉到更加显著的数据特征和模式。

（4）提高计算效率。加快了网络的前向和反向传播过程，从而提高了训练和推理的速度。

2、Dropout层

Dropout的核心思想是在每个训练步骤中，随机“丢弃”神经网络中的一部分神经元，即在前向传播过程中，让这些神经元的输出为0。这意味着在每次训练迭代中，神经网络都会变得“不完整”，因此模型必须适应不同的网络架构。好处如下：

（1）防止过拟合：通过随机丢弃一部分神经元，模型不能依赖于任何一组特定的神经元，因此被迫学习更加鲁棒的特征表示，从而提高了泛化能力。

（2）模型平均：通过训练时使用的是神经网络的一个“子集”，可以看作是从原始网络中采样的一个子模型。在测试时，通过对所有可能的子模型进行平均，可以获得更稳定的预测。实际操作中，这个平均过程通过使用所有神经元并调整它们的激活值来近似实现（通常是乘以保留概率）。

（3）减少相互依存性：Dropout迫使网络中的单个节点不依赖于其他节点的特定配置，从而促进了特征的独立表示。

参数代表舍弃的神经元数量，越高舍弃越多，可根据最终训练结果自行调整。

Using cuda device

Model(
  (conv1): Sequential(
    (0): Conv2d(3, 12, kernel_size=(5, 5), stride=(1, 1))
    (1): BatchNorm2d(12, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (2): ReLU()
  )
  (conv2): Sequential(
    (0): Conv2d(12, 12, kernel_size=(5, 5), stride=(1, 1))
    (1): BatchNorm2d(12, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (2): ReLU()
  )
  (pool3): Sequential(
    (0): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
  )
  (conv4): Sequential(
    (0): Conv2d(12, 24, kernel_size=(5, 5), stride=(1, 1))
    (1): BatchNorm2d(24, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (2): ReLU()
  )
  (conv5): Sequential(
    (0): Conv2d(24, 24, kernel_size=(5, 5), stride=(1, 1))
    (1): BatchNorm2d(24, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (2): ReLU()
  )
  (pool6): Sequential(
    (0): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
  )
  (dropout): Sequential(
    (0): Dropout(p=0.2, inplace=False)
  )
  (fc): Sequential(
    (0): Linear(in_features=60000, out_features=2, bias=True)
  )
)

模型训练

设置动态学习率

def adjust_learning_rate(optimizer, epoch, start_lr):
    # 每 2 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

learn_rate = 1e-4 # 初始学习率
optimizer  = torch.optim.SGD(model.parameters(), lr=learn_rate)

学习率动态下降，同时套用SGD优化器。学习率动态下降可以加速收敛，提高模型性能并防止过拟合。若不知道该如何设定动态规则，也可以直接调用Pytorch中的学习率调度器。

训练函数

# 训练循环
def train(dataloader, model, loss_fn, optimizer):
    size = len(dataloader.dataset)  # 训练集的大小
    num_batches = len(dataloader)   # 批次数目, (size/batch_size，向上取整)

    train_loss, train_acc = 0, 0  # 初始化训练损失和正确率
    
    for X, y in dataloader:  # 获取图片及其标签
        X, y = X.to(device), y.to(device)
        
        # 计算预测误差
        pred = model(X)          # 网络输出
        loss = loss_fn(pred, y)  # 计算网络输出和真实值之间的差距，targets为真实值，计算二者差值即为损失
        
        # 反向传播
        optimizer.zero_grad()  # grad属性归零
        loss.backward()        # 反向传播
        optimizer.step()       # 每一步自动更新
        
        # 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

测试函数

def test (dataloader, model, loss_fn):
    size        = len(dataloader.dataset)  # 测试集的大小
    num_batches = len(dataloader)          # 批次数目, (size/batch_size，向上取整)
    test_loss, test_acc = 0, 0
    
    # 当不进行训练时，停止梯度更新，节省计算内存消耗
    with torch.no_grad():
        for imgs, target in dataloader:
            imgs, target = imgs.to(device), target.to(device)
            
            # 计算loss
            target_pred = model(imgs)
            loss        = loss_fn(target_pred, target)
            
            test_loss += loss.item()
            test_acc  += (target_pred.argmax(1) == target).type(torch.float).sum().item()

    test_acc  /= size
    test_loss /= num_batches

    return test_acc, test_loss

训练和结果可视化

训练结果

loss_fn    = nn.CrossEntropyLoss() # 创建损失函数
epochs     = 40

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

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接口时使用）
    
    model.eval()
    epoch_test_acc, epoch_test_loss = test(test_dl, model, loss_fn)
    
    train_acc.append(epoch_train_acc)
    train_loss.append(epoch_train_loss)
    test_acc.append(epoch_test_acc)
    test_loss.append(epoch_test_loss)
    
    # 获取当前的学习率
    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:50.2%, Train_loss:0.777, Test_acc:50.0%, Test_loss:0.691, Lr:1.00E-04
Epoch: 2, Train_acc:63.5%, Train_loss:0.669, Test_acc:60.5%, Test_loss:0.665, Lr:1.00E-04
Epoch: 3, Train_acc:66.9%, Train_loss:0.630, Test_acc:67.1%, Test_loss:0.630, Lr:9.20E-05
Epoch: 4, Train_acc:72.7%, Train_loss:0.569, Test_acc:65.8%, Test_loss:0.608, Lr:9.20E-05
Epoch: 5, Train_acc:74.5%, Train_loss:0.556, Test_acc:68.4%, Test_loss:0.602, Lr:8.46E-05
Epoch: 6, Train_acc:77.5%, Train_loss:0.523, Test_acc:61.8%, Test_loss:0.650, Lr:8.46E-05
Epoch: 7, Train_acc:75.9%, Train_loss:0.492, Test_acc:73.7%, Test_loss:0.530, Lr:7.79E-05
Epoch: 8, Train_acc:81.1%, Train_loss:0.483, Test_acc:73.7%, Test_loss:0.572, Lr:7.79E-05
Epoch: 9, Train_acc:81.9%, Train_loss:0.460, Test_acc:72.4%, Test_loss:0.532, Lr:7.16E-05
Epoch:10, Train_acc:84.5%, Train_loss:0.421, Test_acc:75.0%, Test_loss:0.552, Lr:7.16E-05
Epoch:11, Train_acc:84.7%, Train_loss:0.428, Test_acc:69.7%, Test_loss:0.557, Lr:6.59E-05
Epoch:12, Train_acc:84.3%, Train_loss:0.413, Test_acc:72.4%, Test_loss:0.492, Lr:6.59E-05
Epoch:13, Train_acc:87.3%, Train_loss:0.396, Test_acc:77.6%, Test_loss:0.487, Lr:6.06E-05
Epoch:14, Train_acc:86.3%, Train_loss:0.384, Test_acc:78.9%, Test_loss:0.511, Lr:6.06E-05
Epoch:15, Train_acc:88.6%, Train_loss:0.379, Test_acc:77.6%, Test_loss:0.483, Lr:5.58E-05
Epoch:16, Train_acc:88.8%, Train_loss:0.361, Test_acc:75.0%, Test_loss:0.538, Lr:5.58E-05
Epoch:17, Train_acc:91.2%, Train_loss:0.341, Test_acc:76.3%, Test_loss:0.502, Lr:5.13E-05
Epoch:18, Train_acc:90.2%, Train_loss:0.351, Test_acc:76.3%, Test_loss:0.463, Lr:5.13E-05
Epoch:19, Train_acc:91.2%, Train_loss:0.340, Test_acc:77.6%, Test_loss:0.465, Lr:4.72E-05
Epoch:20, Train_acc:89.8%, Train_loss:0.340, Test_acc:77.6%, Test_loss:0.504, Lr:4.72E-05
Epoch:21, Train_acc:92.0%, Train_loss:0.329, Test_acc:78.9%, Test_loss:0.461, Lr:4.34E-05
Epoch:22, Train_acc:93.0%, Train_loss:0.325, Test_acc:77.6%, Test_loss:0.435, Lr:4.34E-05
Epoch:23, Train_acc:92.8%, Train_loss:0.318, Test_acc:78.9%, Test_loss:0.461, Lr:4.00E-05
Epoch:24, Train_acc:93.8%, Train_loss:0.316, Test_acc:80.3%, Test_loss:0.487, Lr:4.00E-05
Epoch:25, Train_acc:92.0%, Train_loss:0.320, Test_acc:77.6%, Test_loss:0.497, Lr:3.68E-05
Epoch:26, Train_acc:94.8%, Train_loss:0.300, Test_acc:77.6%, Test_loss:0.435, Lr:3.68E-05
Epoch:27, Train_acc:94.4%, Train_loss:0.300, Test_acc:77.6%, Test_loss:0.436, Lr:3.38E-05
Epoch:28, Train_acc:94.4%, Train_loss:0.297, Test_acc:78.9%, Test_loss:0.500, Lr:3.38E-05
Epoch:29, Train_acc:93.2%, Train_loss:0.304, Test_acc:77.6%, Test_loss:0.467, Lr:3.11E-05
Epoch:30, Train_acc:95.0%, Train_loss:0.278, Test_acc:77.6%, Test_loss:0.467, Lr:3.11E-05
Epoch:31, Train_acc:93.8%, Train_loss:0.293, Test_acc:77.6%, Test_loss:0.477, Lr:2.86E-05
Epoch:32, Train_acc:94.6%, Train_loss:0.290, Test_acc:78.9%, Test_loss:0.446, Lr:2.86E-05
Epoch:33, Train_acc:94.2%, Train_loss:0.286, Test_acc:78.9%, Test_loss:0.432, Lr:2.63E-05
Epoch:34, Train_acc:94.6%, Train_loss:0.270, Test_acc:75.0%, Test_loss:0.475, Lr:2.63E-05
Epoch:35, Train_acc:95.6%, Train_loss:0.268, Test_acc:77.6%, Test_loss:0.460, Lr:2.42E-05
Epoch:36, Train_acc:96.0%, Train_loss:0.261, Test_acc:78.9%, Test_loss:0.463, Lr:2.42E-05
Epoch:37, Train_acc:94.6%, Train_loss:0.269, Test_acc:80.3%, Test_loss:0.461, Lr:2.23E-05
Epoch:38, Train_acc:94.2%, Train_loss:0.268, Test_acc:78.9%, Test_loss:0.405, Lr:2.23E-05
Epoch:39, Train_acc:95.2%, Train_loss:0.269, Test_acc:78.9%, Test_loss:0.432, Lr:2.05E-05
Epoch:40, Train_acc:95.0%, Train_loss:0.264, Test_acc:80.3%, Test_loss:0.449, Lr:2.05E-05
Done

可视化

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

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()

结果评价：

随着训练进程，训练准确性逐渐提高，并趋于稳定。这说明模型在学习数据集的特征，并能越来越好地识别训练数据。测试准确性的提高速度没有训练准确性快，且有较大波动。在整个训练过程中，测试准确性始终低于训练准确性，且训练准确性和测试准确性之间存在一定差异，这些特征表明存在过拟合风险。

指定图片预测

from PIL import Image 

classes = list(train_dataset.class_to_idx)

def predict_one_image(image_path, model, transform, classes):
    
    test_img = Image.open(image_path).convert('RGB')
    # plt.imshow(test_img)  # 展示预测的图片

    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(f'预测结果是：{pred_class}')

# 预测训练集中的某张照片
predict_one_image(image_path='./5-data/test/adidas/1.jpg', 
                  model=model, 
                  transform=train_transforms, 
                  classes=classes)

保存和加载模型

# 模型保存
PATH = './model.pth'  # 保存的参数文件名
torch.save(model.state_dict(), PATH)

# 将参数加载到model当中
model.load_state_dict(torch.load(PATH, map_location=device))

调优尝试

Adam优化器+dropout参数调整

optimizer  = torch.optim.Adam(model.parameters(), lr=learn_rate)

将SGD优化器更换为Adam优化器，同时增加droput参数，从0.2逐步增加到0.5后结果如下

Epoch: 1, Train_acc:53.6%, Train_loss:0.912, Test_acc:51.3%, Test_loss:0.734, Lr:1.00E-04
Epoch: 2, Train_acc:67.1%, Train_loss:0.703, Test_acc:71.1%, Test_loss:0.519, Lr:1.00E-04
Epoch: 3, Train_acc:77.1%, Train_loss:0.467, Test_acc:72.4%, Test_loss:0.655, Lr:9.20E-05
Epoch: 4, Train_acc:83.9%, Train_loss:0.370, Test_acc:80.3%, Test_loss:0.471, Lr:9.20E-05
Epoch: 5, Train_acc:87.6%, Train_loss:0.308, Test_acc:81.6%, Test_loss:0.480, Lr:8.46E-05
Epoch: 6, Train_acc:89.6%, Train_loss:0.259, Test_acc:82.9%, Test_loss:0.408, Lr:8.46E-05
Epoch: 7, Train_acc:94.0%, Train_loss:0.205, Test_acc:81.6%, Test_loss:0.396, Lr:7.79E-05
Epoch: 8, Train_acc:96.0%, Train_loss:0.156, Test_acc:82.9%, Test_loss:0.349, Lr:7.79E-05
Epoch: 9, Train_acc:96.6%, Train_loss:0.160, Test_acc:81.6%, Test_loss:0.425, Lr:7.16E-05
Epoch:10, Train_acc:96.2%, Train_loss:0.145, Test_acc:85.5%, Test_loss:0.363, Lr:7.16E-05
Epoch:11, Train_acc:97.6%, Train_loss:0.123, Test_acc:82.9%, Test_loss:0.383, Lr:6.59E-05
Epoch:12, Train_acc:99.2%, Train_loss:0.092, Test_acc:86.8%, Test_loss:0.369, Lr:6.59E-05
Epoch:13, Train_acc:99.0%, Train_loss:0.088, Test_acc:85.5%, Test_loss:0.349, Lr:6.06E-05
Epoch:14, Train_acc:99.0%, Train_loss:0.076, Test_acc:85.5%, Test_loss:0.395, Lr:6.06E-05
Epoch:15, Train_acc:99.2%, Train_loss:0.074, Test_acc:82.9%, Test_loss:0.375, Lr:5.58E-05
Epoch:16, Train_acc:100.0%, Train_loss:0.063, Test_acc:84.2%, Test_loss:0.419, Lr:5.58E-05
Epoch:17, Train_acc:100.0%, Train_loss:0.053, Test_acc:82.9%, Test_loss:0.469, Lr:5.13E-05
Epoch:18, Train_acc:99.4%, Train_loss:0.061, Test_acc:85.5%, Test_loss:0.319, Lr:5.13E-05
Epoch:19, Train_acc:99.2%, Train_loss:0.063, Test_acc:85.5%, Test_loss:0.400, Lr:4.72E-05
Epoch:20, Train_acc:100.0%, Train_loss:0.050, Test_acc:81.6%, Test_loss:0.364, Lr:4.72E-05
Epoch:21, Train_acc:100.0%, Train_loss:0.046, Test_acc:82.9%, Test_loss:0.397, Lr:4.34E-05
Epoch:22, Train_acc:100.0%, Train_loss:0.043, Test_acc:81.6%, Test_loss:0.440, Lr:4.34E-05
Epoch:23, Train_acc:99.8%, Train_loss:0.041, Test_acc:82.9%, Test_loss:0.348, Lr:4.00E-05
Epoch:24, Train_acc:99.8%, Train_loss:0.041, Test_acc:80.3%, Test_loss:0.413, Lr:4.00E-05
Epoch:25, Train_acc:100.0%, Train_loss:0.038, Test_acc:82.9%, Test_loss:0.420, Lr:3.68E-05
Epoch:26, Train_acc:100.0%, Train_loss:0.036, Test_acc:80.3%, Test_loss:0.345, Lr:3.68E-05
Epoch:27, Train_acc:100.0%, Train_loss:0.040, Test_acc:81.6%, Test_loss:0.440, Lr:3.38E-05
Epoch:28, Train_acc:99.6%, Train_loss:0.043, Test_acc:82.9%, Test_loss:0.380, Lr:3.38E-05
Epoch:29, Train_acc:99.8%, Train_loss:0.034, Test_acc:84.2%, Test_loss:0.368, Lr:3.11E-05
Epoch:30, Train_acc:100.0%, Train_loss:0.033, Test_acc:82.9%, Test_loss:0.415, Lr:3.11E-05
Epoch:31, Train_acc:100.0%, Train_loss:0.029, Test_acc:81.6%, Test_loss:0.459, Lr:2.86E-05
Epoch:32, Train_acc:100.0%, Train_loss:0.030, Test_acc:81.6%, Test_loss:0.340, Lr:2.86E-05
Epoch:33, Train_acc:100.0%, Train_loss:0.031, Test_acc:82.9%, Test_loss:0.355, Lr:2.63E-05
Epoch:34, Train_acc:100.0%, Train_loss:0.028, Test_acc:81.6%, Test_loss:0.454, Lr:2.63E-05
Epoch:35, Train_acc:100.0%, Train_loss:0.026, Test_acc:80.3%, Test_loss:0.349, Lr:2.42E-05
Epoch:36, Train_acc:100.0%, Train_loss:0.025, Test_acc:84.2%, Test_loss:0.425, Lr:2.42E-05
Epoch:37, Train_acc:99.8%, Train_loss:0.029, Test_acc:84.2%, Test_loss:0.354, Lr:2.23E-05
Epoch:38, Train_acc:100.0%, Train_loss:0.025, Test_acc:81.6%, Test_loss:0.439, Lr:2.23E-05
Epoch:39, Train_acc:100.0%, Train_loss:0.025, Test_acc:82.9%, Test_loss:0.378, Lr:2.05E-05
Epoch:40, Train_acc:100.0%, Train_loss:0.023, Test_acc:84.2%, Test_loss:0.427, Lr:2.05E-05
Done

可以看到，测试集准确率有所提高但提升幅度不大，同时过拟合特征依然存在。并且也不能确定是因为变更比例导致的效果提升，多次运行效果呈现不同。

其他调优尝试

后续又尝试了调整卷积核大小，将学习率由动态改为静态，准确率都未能得到明显提升。后续有机会可再多做尝试，先争取在不改变CNN架构情况下尝试能否改善准确率，若效果不好则可以考虑调整模型架构。

XiaoLaJi2019

关注

31
点赞
踩
19

收藏

觉得还不错? 一键收藏
1
评论
深度学习P5：运动鞋识别

用Pytorch分析图片的常规预处理，注释掉的翻转其实是训练数据前的常见操作，一些翻转、裁剪等可能更有利于模型学习图片特征，根据具体情况可尝试是否能提升模型的准确率。同时，稀疏激活意味着网络能够在较少的特征上集中学习，这可以帮助模型捕捉到更加显著的数据特征和模式。Dropout的核心思想是在每个训练步骤中，随机“丢弃”神经网络中的一部分神经元，即在前向传播过程中，让这些神经元的输出为0。在前面的CNN网络中并没有设计该激活函数，该函数的主要作用是通过引入非线性，提高了神经网络模型的学习能力和表达能力。
复制链接

扫一扫