微调ResNet18 on MNIST精度99.59%以上

最新推荐文章于 2024-06-08 16:29:58 发布
最新推荐文章于 2024-06-08 16:29:58 发布
阅读量2.3k 收藏 23
点赞数 4
分类专栏： 笔记 文章标签： pytorch 深度学习 神经网络 机器学习
版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。
本文链接：https://blog.csdn.net/qq_41249412/article/details/115533857
版权
时间20210408
作者：知道许多的橘子
实现：Resnet18对MNIST数据集的分类，作为一个入门
测试集准确度：99.59%
实现框架pytorch
数据增强方法：Normalize
训练次数：60
阶段学习率[0-20]：0.001||[20-40]：0.0003||[40-60]：0.0001
优化器：optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
# In[1] 导入所需工具包
import torch
import torch.nn as nn
import torchvision
from torchvision import datasets,transforms
import time
from torch.nn import functional as F
from math import floor, ceil
#import torchvision.transforms as transforms
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(device)
# In[1] 设置超参数
num_epochs = 60
batch_size = 100
learning_rate = 0.001

# In[2] 获取数据包括训练数据和测试数据
train_loader =torch.utils.data.DataLoader(
        torchvision.datasets.MNIST('MNIST',train=True,download=True,
                                   transform=torchvision.transforms.Compose([
                                           torchvision.transforms.ToTensor(),
                                           torchvision.transforms.Normalize(
                                                   (0.1307,),(0.3081,))
                                           ])),
        batch_size=batch_size,shuffle=True)

test_loader =torch.utils.data.DataLoader(
        torchvision.datasets.MNIST('MNIST',train=False,download=True,
                                   transform=torchvision.transforms.Compose([
                                           torchvision.transforms.ToTensor(),
                                           torchvision.transforms.Normalize(
                                                   (0.1307,),(0.3081,))
                                           ])),
        batch_size=batch_size,shuffle=True)
# In[1] 定义卷积核
def conv3x3(in_channels, out_channels, stride=1):
    return nn.Conv2d(in_channels, out_channels, kernel_size=3,
                     stride=stride, padding=1, bias=True)

# In[1] 定义残差块
class ResidualBlock(nn.Module):
    def __init__(self, in_channels, out_channels, stride=1, downsample=None):
        super(ResidualBlock, self).__init__()
        self.conv1 = conv3x3(in_channels, out_channels, stride)
        self.bn1 = nn.BatchNorm2d(out_channels)
        self.relu = nn.ReLU(inplace=True)
        self.conv2 = conv3x3(out_channels, out_channels)
        self.bn2 = nn.BatchNorm2d(out_channels)
        self.downsample = downsample

    def forward(self, x):
        residual = x
        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)
        out = self.conv2(out)
        out = self.bn2(out)
        # 下采样
        if self.downsample:
            residual = self.downsample(x)
        out += residual
        out = self.relu(out)
        return out
# In[1] 搭建残差神经网络
class ResNet(nn.Module):
    def __init__(self, block, layers, num_classes=10):
        super(ResNet, self).__init__()
        self.in_channels = 16
        self.conv = conv3x3(1, 16)
        self.bn = nn.BatchNorm2d(16)
        self.relu = nn.ReLU(inplace=True)
        # 构建残差块,恒等映射
        # in_channels == out_channels and stride = 1 所以这里我们构建残差块,没有下采样
        self.layer1 = self.make_layer(block, 16, layers[0], stride=1)
        # 不构建残差块,进行了下采样
        # layers中记录的是数字,表示对应位置的残差块数目
        self.layer2 = self.make_layer(block, 32, layers[1], 2)
        # 不构建残差块,进行了下采样
        self.layer3 = self.make_layer(block, 64, layers[2], 2)
        self.avg_pool = nn.AvgPool2d(8)
        self.fc1 = nn.Linear(3136, 128)
        self.normfc12 = nn.LayerNorm((128), eps=1e-5)
        self.fc2 = nn.Linear(128, num_classes)

    def make_layer(self, block, out_channels, blocks, stride=1):
        downsample = None
        if (stride != 1) or (self.in_channels != out_channels):
            downsample = nn.Sequential(
                conv3x3(self.in_channels, out_channels, stride=stride),
                nn.BatchNorm2d(out_channels))
        layers = []
        layers.append(block(self.in_channels, out_channels, stride, downsample))
        # 当out_channels = 32时,in_channels也变成32了
        self.in_channels = out_channels
        # blocks是残差块的数目
        # 残差块之后的网络结构,是out_channels->out_channels的
        # 可以说,make_layer做的是输出尺寸相同的所有网络结构
        # 由于输出尺寸会改变,我们用make_layers去生成一大块对应尺寸完整网络结构
        for i in range(1, blocks):
            layers.append(block(out_channels, out_channels))
        return nn.Sequential(*layers)

    def forward(self, x):
        out = self.conv(x)
        out = self.bn(out)
        out = self.relu(out)
        # layer1是三块in_channels等于16的网络结构，包括三个恒等映射
        out = self.layer1(out)
        # layer2包括了16->32下采样,然后是32的三个恒等映射
        out = self.layer2(out)
        # layer3包括了32->64的下采样,然后是64的三个恒等映射
        out = self.layer3(out)
        #out = self.avg_pool(out)
        # 全连接压缩
        # out.size(0)可以看作是batch_size
        out = out.view(out.size(0), -1)
        out = self.fc1(out)
        out = self.normfc12(out)
        out = self.relu(out)
        out = self.fc2(out)
        return out
# In[1] 定义模型和损失函数
# [2,2,2]表示的是不同in_channels下的恒等映射数目
model = ResNet(ResidualBlock, [2, 2, 2]).to(device)
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
# In[1] 设置一个通过优化器更新学习率的函数
def update_lr(optimizer, lr):
    for param_group in optimizer.param_groups:
        param_group['lr'] = lr
# In[1] 定义测试函数
def test(model,test_loader):
    
    model.eval()
    with torch.no_grad():
        correct = 0
        total = 0
        for images, labels in test_loader:
            images = images.to(device)
            labels = labels.to(device)
            outputs = model(images)
            _, predicted = torch.max(outputs.data, 1)
            total += labels.size(0)
            correct += (predicted == labels).sum().item()
    
        print('Accuracy of the model on the test images: {} %'.format(100 * correct / total))

        
# In[1] 训练模型更新学习率
total_step = len(train_loader)
curr_lr = learning_rate
for epoch in range(num_epochs):
    in_epoch = time.time()
    for i, (images, labels) in enumerate(train_loader):
        images = images.to(device)
        labels = labels.to(device)

        # Forward pass
        outputs = model(images)
        loss = criterion(outputs, labels)

        # Backward and optimize
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

        if (i + 1) % 100 == 0:
            print("Epoch [{}/{}], Step [{}/{}] Loss: {:.4f}"
                  .format(epoch + 1, num_epochs, i + 1, total_step, loss.item()))
    test(model,test_loader)
    out_epoch = time.time()
    print(f"use {(out_epoch-in_epoch)//60}min{(out_epoch-in_epoch)%60}s")
    if (epoch + 1) % 20 == 0:
        curr_lr /= 3
        update_lr(optimizer, curr_lr)
# In[1] 测试模型并保存
model.eval()
with torch.no_grad():
    correct = 0
    total = 0
    for images, labels in test_loader:
        images = images.to(device)
        labels = labels.to(device)
        outputs = model(images)
        _, predicted = torch.max(outputs.data, 1)
        total += labels.size(0)
        correct += (predicted == labels).sum().item()

    print('Accuracy of the model on the test images: {} %'.format(100 * correct / total))

torch.save(model.state_dict(), 'resnet.ckpt')

cuda
Epoch [1/60], Step [100/600] Loss: 0.0817
Epoch [1/60], Step [200/600] Loss: 0.0185
Epoch [1/60], Step [300/600] Loss: 0.0325
Epoch [1/60], Step [400/600] Loss: 0.1492
Epoch [1/60], Step [500/600] Loss: 0.1131
Epoch [1/60], Step [600/600] Loss: 0.0087
Accuracy of the model on the test images: 99.03 %
use 0.0min18.335710287094116s
Epoch [2/60], Step [100/600] Loss: 0.0504
Epoch [2/60], Step [200/600] Loss: 0.1095
Epoch [2/60], Step [300/600] Loss: 0.1329
Epoch [2/60], Step [400/600] Loss: 0.0889
Epoch [2/60], Step [500/600] Loss: 0.0196
Epoch [2/60], Step [600/600] Loss: 0.0263
Accuracy of the model on the test images: 99.05 %
use 0.0min17.68439769744873s
Epoch [3/60], Step [100/600] Loss: 0.0244
Epoch [3/60], Step [200/600] Loss: 0.0551
Epoch [3/60], Step [300/600] Loss: 0.0242
Epoch [3/60], Step [400/600] Loss: 0.0272
Epoch [3/60], Step [500/600] Loss: 0.0049
Epoch [3/60], Step [600/600] Loss: 0.0472
Accuracy of the model on the test images: 99.08 %
use 0.0min17.7007839679718s
Epoch [4/60], Step [100/600] Loss: 0.0346
Epoch [4/60], Step [200/600] Loss: 0.0120
Epoch [4/60], Step [300/600] Loss: 0.0021
Epoch [4/60], Step [400/600] Loss: 0.0050
Epoch [4/60], Step [500/600] Loss: 0.0097
Epoch [4/60], Step [600/600] Loss: 0.0522
Accuracy of the model on the test images: 99.3 %
use 0.0min17.670555114746094s
Epoch [5/60], Step [100/600] Loss: 0.0028
Epoch [5/60], Step [200/600] Loss: 0.0111
Epoch [5/60], Step [300/600] Loss: 0.0108
Epoch [5/60], Step [400/600] Loss: 0.0018
Epoch [5/60], Step [500/600] Loss: 0.0610
Epoch [5/60], Step [600/600] Loss: 0.0345
Accuracy of the model on the test images: 99.26 %
use 0.0min17.79431986808777s
Epoch [6/60], Step [100/600] Loss: 0.0010
Epoch [6/60], Step [200/600] Loss: 0.0095
Epoch [6/60], Step [300/600] Loss: 0.0013
Epoch [6/60], Step [400/600] Loss: 0.0085
Epoch [6/60], Step [500/600] Loss: 0.0027
Epoch [6/60], Step [600/600] Loss: 0.0252
Accuracy of the model on the test images: 99.06 %
use 0.0min17.835293769836426s
Epoch [7/60], Step [100/600] Loss: 0.0406
Epoch [7/60], Step [200/600] Loss: 0.0022
Epoch [7/60], Step [300/600] Loss: 0.0005
Epoch [7/60], Step [400/600] Loss: 0.0031
Epoch [7/60], Step [500/600] Loss: 0.0008
Epoch [7/60], Step [600/600] Loss: 0.0170
Accuracy of the model on the test images: 99.24 %
use 0.0min17.662654638290405s
Epoch [8/60], Step [100/600] Loss: 0.0009
Epoch [8/60], Step [200/600] Loss: 0.0026
Epoch [8/60], Step [300/600] Loss: 0.0040
Epoch [8/60], Step [400/600] Loss: 0.0008
Epoch [8/60], Step [500/600] Loss: 0.0005
Epoch [8/60], Step [600/600] Loss: 0.0024
Accuracy of the model on the test images: 98.71 %
use 0.0min17.942986488342285s
Epoch [9/60], Step [100/600] Loss: 0.0005
Epoch [9/60], Step [200/600] Loss: 0.0008
Epoch [9/60], Step [300/600] Loss: 0.0010
Epoch [9/60], Step [400/600] Loss: 0.0200
Epoch [9/60], Step [500/600] Loss: 0.0171
Epoch [9/60], Step [600/600] Loss: 0.0024
Accuracy of the model on the test images: 99.23 %
use 0.0min17.72422456741333s
Epoch [10/60], Step [100/600] Loss: 0.0007
Epoch [10/60], Step [200/600] Loss: 0.0084
Epoch [10/60], Step [300/600] Loss: 0.0002
Epoch [10/60], Step [400/600] Loss: 0.0007
Epoch [10/60], Step [500/600] Loss: 0.0007
Epoch [10/60], Step [600/600] Loss: 0.0229
Accuracy of the model on the test images: 99.15 %
use 0.0min17.694152116775513s
Epoch [11/60], Step [100/600] Loss: 0.0103
Epoch [11/60], Step [200/600] Loss: 0.0186
Epoch [11/60], Step [300/600] Loss: 0.0355
Epoch [11/60], Step [400/600] Loss: 0.0014
Epoch [11/60], Step [500/600] Loss: 0.0118
Epoch [11/60], Step [600/600] Loss: 0.0038
Accuracy of the model on the test images: 99.4 %
use 0.0min17.857590675354004s
Epoch [12/60], Step [100/600] Loss: 0.0001
Epoch [12/60], Step [200/600] Loss: 0.0001
Epoch [12/60], Step [300/600] Loss: 0.0008
Epoch [12/60], Step [400/600] Loss: 0.0001
Epoch [12/60], Step [500/600] Loss: 0.0007
Epoch [12/60], Step [600/600] Loss: 0.0005
Accuracy of the model on the test images: 99.25 %
use 0.0min17.56403350830078s
Epoch [13/60], Step [100/600] Loss: 0.0018
Epoch [13/60], Step [200/600] Loss: 0.0809
Epoch [13/60], Step [300/600] Loss: 0.0101
Epoch [13/60], Step [400/600] Loss: 0.0036
Epoch [13/60], Step [500/600] Loss: 0.0009
Epoch [13/60], Step [600/600] Loss: 0.0182
Accuracy of the model on the test images: 99.19 %
use 0.0min17.759077072143555s
Epoch [14/60], Step [100/600] Loss: 0.0014
Epoch [14/60], Step [200/600] Loss: 0.0371
Epoch [14/60], Step [300/600] Loss: 0.0003
Epoch [14/60], Step [400/600] Loss: 0.0143
Epoch [14/60], Step [500/600] Loss: 0.0002
Epoch [14/60], Step [600/600] Loss: 0.0371
Accuracy of the model on the test images: 99.17 %
use 0.0min17.763488292694092s
Epoch [15/60], Step [100/600] Loss: 0.0002
Epoch [15/60], Step [200/600] Loss: 0.0002
Epoch [15/60], Step [300/600] Loss: 0.0000
Epoch [15/60], Step [400/600] Loss: 0.0001
Epoch [15/60], Step [500/600] Loss: 0.0000
Epoch [15/60], Step [600/600] Loss: 0.0018
Accuracy of the model on the test images: 99.26 %
use 0.0min17.66961097717285s
Epoch [16/60], Step [100/600] Loss: 0.0002
Epoch [16/60], Step [200/600] Loss: 0.0001
Epoch [16/60], Step [300/600] Loss: 0.0394
Epoch [16/60], Step [400/600] Loss: 0.0123
Epoch [16/60], Step [500/600] Loss: 0.0000
Epoch [16/60], Step [600/600] Loss: 0.0323
Accuracy of the model on the test images: 99.33 %
use 0.0min17.823559761047363s
Epoch [17/60], Step [100/600] Loss: 0.0019
Epoch [17/60], Step [200/600] Loss: 0.0004
Epoch [17/60], Step [300/600] Loss: 0.0005
Epoch [17/60], Step [400/600] Loss: 0.0003
Epoch [17/60], Step [500/600] Loss: 0.0004
Epoch [17/60], Step [600/600] Loss: 0.0014
Accuracy of the model on the test images: 99.4 %
use 0.0min17.818267107009888s
Epoch [18/60], Step [100/600] Loss: 0.0026
Epoch [18/60], Step [200/600] Loss: 0.0069
Epoch [18/60], Step [300/600] Loss: 0.0042
Epoch [18/60], Step [400/600] Loss: 0.0014
Epoch [18/60], Step [500/600] Loss: 0.0010
Epoch [18/60], Step [600/600] Loss: 0.0003
Accuracy of the model on the test images: 99.39 %
use 0.0min17.786447763442993s
Epoch [19/60], Step [100/600] Loss: 0.0001
Epoch [19/60], Step [200/600] Loss: 0.0002
Epoch [19/60], Step [300/600] Loss: 0.0001
Epoch [19/60], Step [400/600] Loss: 0.0001
Epoch [19/60], Step [500/600] Loss: 0.0384
Epoch [19/60], Step [600/600] Loss: 0.0019
Accuracy of the model on the test images: 99.32 %
use 0.0min17.744271516799927s
Epoch [20/60], Step [100/600] Loss: 0.0050
Epoch [20/60], Step [200/600] Loss: 0.0001
Epoch [20/60], Step [300/600] Loss: 0.0023
Epoch [20/60], Step [400/600] Loss: 0.0000
Epoch [20/60], Step [500/600] Loss: 0.0034
Epoch [20/60], Step [600/600] Loss: 0.0008
Accuracy of the model on the test images: 99.35 %
use 0.0min17.80747652053833s
Epoch [21/60], Step [100/600] Loss: 0.0001
Epoch [21/60], Step [200/600] Loss: 0.0001
Epoch [21/60], Step [300/600] Loss: 0.0003
Epoch [21/60], Step [400/600] Loss: 0.0000
Epoch [21/60], Step [500/600] Loss: 0.0003
Epoch [21/60], Step [600/600] Loss: 0.0000
Accuracy of the model on the test images: 99.46 %
use 0.0min17.777599573135376s
Epoch [22/60], Step [100/600] Loss: 0.0003
Epoch [22/60], Step [200/600] Loss: 0.0001
Epoch [22/60], Step [300/600] Loss: 0.0000
Epoch [22/60], Step [400/600] Loss: 0.0000
Epoch [22/60], Step [500/600] Loss: 0.0003
Epoch [22/60], Step [600/600] Loss: 0.0000
Accuracy of the model on the test images: 99.55 %
use 0.0min17.79033088684082s
Epoch [23/60], Step [100/600] Loss: 0.0000
Epoch [23/60], Step [200/600] Loss: 0.0002
Epoch [23/60], Step [300/600] Loss: 0.0000
Epoch [23/60], Step [400/600] Loss: 0.0000
Epoch [23/60], Step [500/600] Loss: 0.0000
Epoch [23/60], Step [600/600] Loss: 0.0000
Accuracy of the model on the test images: 99.59 %
use 0.0min17.915147304534912s
Epoch [24/60], Step [100/600] Loss: 0.0001
Epoch [24/60], Step [200/600] Loss: 0.0000
Epoch [24/60], Step [300/600] Loss: 0.0000
Epoch [24/60], Step [400/600] Loss: 0.0000
Epoch [24/60], Step [500/600] Loss: 0.0000
Epoch [24/60], Step [600/600] Loss: 0.0000
Accuracy of the model on the test images: 99.57 %
use 0.0min17.72119116783142s
Epoch [25/60], Step [100/600] Loss: 0.0001
Epoch [25/60], Step [200/600] Loss: 0.0001
Epoch [25/60], Step [300/600] Loss: 0.0001
Epoch [25/60], Step [400/600] Loss: 0.0000
Epoch [25/60], Step [500/600] Loss: 0.0000
Epoch [25/60], Step [600/600] Loss: 0.0000
Accuracy of the model on the test images: 99.57 %
use 0.0min17.888681411743164s
Epoch [26/60], Step [100/600] Loss: 0.0000
Epoch [26/60], Step [200/600] Loss: 0.0000
Epoch [26/60], Step [300/600] Loss: 0.0000
Epoch [26/60], Step [400/600] Loss: 0.0000
Epoch [26/60], Step [500/600] Loss: 0.0000
Epoch [26/60], Step [600/600] Loss: 0.0000
Accuracy of the model on the test images: 99.56 %
use 0.0min17.78260850906372s
Epoch [27/60], Step [100/600] Loss: 0.0000
Epoch [27/60], Step [200/600] Loss: 0.0000
Epoch [27/60], Step [300/600] Loss: 0.0000
Epoch [27/60], Step [400/600] Loss: 0.0000
Epoch [27/60], Step [500/600] Loss: 0.0000
Epoch [27/60], Step [600/600] Loss: 0.0001
Accuracy of the model on the test images: 99.57 %
use 0.0min17.75735902786255s
Epoch [28/60], Step [100/600] Loss: 0.0000
Epoch [28/60], Step [200/600] Loss: 0.0000
Epoch [28/60], Step [300/600] Loss: 0.0000
Epoch [28/60], Step [400/600] Loss: 0.0000
Epoch [28/60], Step [500/600] Loss: 0.0000
Epoch [28/60], Step [600/600] Loss: 0.0000
Accuracy of the model on the test images: 99.57 %
use 0.0min17.589683771133423s
Epoch [29/60], Step [100/600] Loss: 0.0000
Epoch [29/60], Step [200/600] Loss: 0.0000
Epoch [29/60], Step [300/600] Loss: 0.0000
Epoch [29/60], Step [400/600] Loss: 0.0000
Epoch [29/60], Step [500/600] Loss: 0.0000
Epoch [29/60], Step [600/600] Loss: 0.0000
Accuracy of the model on the test images: 99.57 %
use 0.0min17.815569400787354s
Epoch [30/60], Step [100/600] Loss: 0.0000
Epoch [30/60], Step [200/600] Loss: 0.0000
Epoch [30/60], Step [300/600] Loss: 0.0000
Fu_Xingwen
关注
  • 4
    点赞
  • 踩
  • 23
    收藏
    觉得还不错? 一键收藏
  • 0
    评论
专栏目录
从零开始大模型开发与微调：ResNet实战
禅与计算机程序设计艺术
05-08 736
从零开始大模型开发与微调：ResNet实战 作者：禅与计算机程序设计艺术 1. 背景介绍 1.1 深度学习的发展历程 1.1.1 人工智能的起源与发展
【Pytorch】ResNet-18实现Cifar-10图像分类
热门推荐
AMG_GT的博客
04-23 2万+
一、ResNet-18网络结构 ResNet全名Residual Network残差网络。Kaiming He 的《Deep Residual Learning for Image Recognition》获得了CVPR最佳论文。他提出的深度残差网络在2015年可以说是洗刷了图像方面的各大比赛，以绝对优势取得了多个比赛的冠军。而且它在保证网络精度的前提下，将网络的深度达到了1...
使用TensorFlow和Keras对以ResNet50模型进行微调
最新发布
heeheeai的博客
06-08 1101
以下是使用ResNet50进行微调以识别特定的新东西的代码演示。将使用TensorFlow和Keras进行这个任务。数据集下载地址，解压到工程里面去： 原始代码：​ ``` 解析 安装必要的库： 导入库： 加载ResNet50模型，并去掉顶层： 添加自定义顶层： 冻结base_model的所有卷积层： 编译模型： 准备数据： 假设数据存储在和中，并且每个类有一个文件夹。 训练模型： 解冻部分或全部的卷积层并继续训练： 在这段代码中，“添加自定义顶层”是指在预训练的R
用Pytorch自建6层神经网络训练Fashion-MNIST数据集，测试准确率达到 92%
frank 的专栏
09-10 1万+
一般的深度学习入门例子是 MNIST 的训练和测试，几乎就算是深度学习领域的 HELLO WORLD 了，但是，有一个问题是，MNIST 太简单了，初学者闭着眼镜随便构造几层网络就可以将准确率提升到 90% 以上。但是，初学者这算入门了吗？ 答案是没有。 现实开发当中的例子可没有这么简单，如果让初学者直接去上手 VOC 或者是 COCO 这样的数据集，很可能自己搭建的神经网络准确率不超过 30%。...
mxnet微调resnet-18
斯人若彩虹，遇上方知有！
08-25 1407
本文的全部思想来自李沐大神的B站课程！有兴趣的可以去学习！！！ 1、导入相关的包 #导入实验所需要的包 import d2lzh as d2l from mxnet import gluon,init,nd from mxnet.gluon import data as gdata,loss as gloss,model_zoo from mxnet.gluon import utils a...
Pytorch之模型微调（Finetune）——用Resnet18进行蚂蚁蜜蜂二分类为例
垃圾就应该待在垃圾桶里的博客
09-02 3602
Pytorch之模型微调（Finetune）——手写数字集为例 文章目录Pytorch之模型微调（Finetune）——手写数字集为例前言一、Transfer Learning and Model Finetune1.Transform Learning（迁移学习）2.Model Finetune（模型微调）二、Pytorch中的Finetune1.模型微调的步骤2.模型微调训练方法3.Pytorch——基于Resnet18蚂蚁蜜蜂二分类源码参考 前言 模型微调在迁移学习中（Transfer Learn
微调vgg16预训练模型
WonderKing'blog
05-18 1261
微调代码 只训练全连接层 model = torch.load( '../model/20220509-pretrain-vgg16-数据增强-5e-05.pth') num_ftrs = model.classifier[6].in_features model.classifier[6] = nn.Linear(num_ftrs, 8) for name, child in model.named_children(): if name in ['avgpool','classi
对Lenet 神经网络进行微调 ，使其适应 Mnist 数据集
Andrew的博客
03-08 267
对Lenet 神经网络进行微调 ，使其适应 Mnist 数据集 Lenet 神经网络的结构图及特征提取过程如下所示： 由于 Mnist 数据集中图片大小为 28281 的灰度图片，而 Lenet 神经网络的输入 为 32321，故需要对 Lenet 神经网络进行微调。 Lenet 进行微调后的结构如下所示： 线理解原理，再看代码。 mnist_lenet5_forward.py #codi...
resnet18-5c106cde.pth
01-08
标题中的"resnet18-5c106cde.pth"是一个预训练模型的权重文件，这在深度学习领域非常常见。ResNet18是Residual Network（残差网络）系列模型的一个变体，由Kaiming He等人在2015年提出。这个模型因其18层的深度而...
resnet18-f37072fd.pth
11-20
resnet18-f37072fd.pth
resnet18-5c106cde.pth和resnet101-5d3b4d8f.pth
08-22
总之，“resnet18-5c106cde.pth”和“resnet101-5d3b4d8f.pth”是PyTorch中ResNet18和ResNet101模型的预训练权重，它们在深度学习领域有着广泛的应用，尤其是在图像识别、物体检测和语义分割等任务中，极大地提高了...
ResNet50_fashion_mnist.py
04-27
用keras实现resnet50,用的数据集是fashion_mnist 实现后准确率是91.3%，损失率是53.7%
resnet18-train-test.py
06-12
resnet18_train_test.py
【pytorch】nni微调pytorch mnist
luolinll1212的专栏
11-21 647
1，文件配置 rose@rose-machine:~/pytorch_c++/test$ tree . ├── config.yml # 微调配置文件 ├── mnist.py # 启动文件 └── search_space.json # 参数配置文件 2，启动与停止 nnictl create --config c...
神经网络--基于mnist数据集取得最高的识别准确率
02-01 3275
今天来学习一下如何基于mnist数据集取得最高的识别准确率，本文是从零开始的，如有需要可自行跳至所需内容
keras模拟alexnet， mnist test（第三次）（改变dropout的权值)（最高精度99.51）
weixin_42160791的博客
03-25 267
from keras.models import Sequential from keras.layers import Input, Dense, Dropout, Activation, BatchNormalization from keras.models import Model from keras.optimizers import SGD from keras.dataset...
如何快速高效的训练ResNet，各种奇技淫巧（五）：超参数
AI公园
08-26 9294
点击上方“AI公园”，关注公众号，选择加“星标“或“置顶”作者：David Page编译：ronghuaiyang导读这个系列介绍了如何在CIFAR10上高效的训练Res...
经典神经网络(6)ResNet及其在Fashion-MNIST数据集上的应用
qq_44665283的博客
05-31 1718
经典神经网络(6)ResNet及其在Fashion-MNIST数据集上的应用
【联邦学习实战】基于FATE框架的MNIST手写数字识别——卷积神经网络
HERODING23的博客
05-01 4289
在FATE上基于mnist数据集训练手写数字识别模型前言1. 下载MNIST数据集1.1 KagglePython格式转换2. 数据集分割3.数据集上传3.1 docker上传3.2 rz工具3.3 FATE数据上传参考链接 前言 ATE是微众银行开发的联邦学习平台，是全球首个工业级的联邦学习开源框架，在github上拥有超过4000stars，可谓是相当有名气的，该平台为联邦学习提供了完整的生态和社区支持，为联邦学习初学者提供了很好的环境，否则利用python从零开发，那将会是一件非常痛苦的事情。本篇博客
resnet18训练mnist数据集
09-15
要使用ResNet-18来训练MNIST数据集，首先需要导入相应的库，并获取MNIST数据集。然后，需要调整ResNet-18模型的全连接层的输出，以适应MNIST数据集的10个类别。接下来，可以开启训练过程，并在训练过程中显示损失值。训练完成后，可以保存训练好的模型权重文件。最后，在测试集上测试训练后模型的准确率。 下面是一种可能的实现： ```python import torch import torch.nn as nn import torch.optim as optim from torchvision import datasets, models, transforms # 导入MNIST数据集 train_dataset = datasets.MNIST(root='./data', train=True, transform=transforms.ToTensor(), download=True) test_dataset = datasets.MNIST(root='./data', train=False, transform=transforms.ToTensor(), download=True) # 创建ResNet-18模型 model = models.resnet18(pretrained=True) in_features = model.fc.in_features model.fc = nn.Linear(in_features, 10) # 设置优化器和损失函数 optimizer = optim.SGD(model.parameters(), lr=0.001, momentum=0.9) criterion = nn.CrossEntropyLoss() # 开始训练 num_epochs = 10 batch_size = 64 train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=batch_size, shuffle=True) for epoch in range(num_epochs): running_loss = 0.0 for images, labels in train_loader: optimizer.zero_grad() outputs = model(images) loss = criterion(outputs, labels) loss.backward() optimizer.step() running_loss += loss.item() print("Epoch {} - Training loss: {:.4f}".format(epoch+1, running_loss / len(train_loader))) # 保存训练好的模型权重文件 torch.save(model.state_dict(), 'resnet18_mnist.pth') # 在测试集上测试模型准确率 test_loader = torch.utils.data.DataLoader(test_dataset, batch_size=batch_size, shuffle=False) correct = 0 total = 0 with torch.no_grad(): for images, labels in test_loader: outputs = model(images) _, predicted = torch.max(outputs.data, 1) total += labels.size(0) correct += (predicted == labels).sum().item() accuracy = 100 * correct / total print("Test accuracy: {:.2f}%".format(accuracy)) ```

“相关推荐”对你有帮助么？

  • 非常没帮助
  • 没帮助
  • 一般
  • 有帮助
  • 非常有帮助
提交
评论
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付元
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明：

1.余额是钱包充值的虚拟货币，按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载，可以购买VIP、付费专栏及课程。

余额充值