完整代码位置https://gitee.com/chuge325/base_machinelearning.git
这里还参考了tensorflow的官方文档

但是由于是pytorch训练的差别还是比较大的，经过多次尝试完成了训练
硬件是两张v100

1超参数训练代码

这个代码可以查看每次训练的loss曲线和超参数的对比信息

import pandas as pd
import torch
from utils.DataLoader import MNIST_data
from torchvision import transforms
from utils.DataLoader import RandomRotation,RandomShift
from model.AlexNet import Net
# from model.Net import Net
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.optim import lr_scheduler
from torch.autograd import Variable
from torch.utils.tensorboard import SummaryWriter
from tensorboard.plugins.hparams import api as hp

from sklearn.model_selection import ParameterSampler
train_df = pd.read_csv('../datasets/digit-recognizer/train.csv')
n_train = len(train_df)
n_pixels = len(train_df.columns) - 1
n_class = len(set(train_df['label']))

# 定义超参数搜索空间
HP_LEARNING_RATE = hp.HParam('learning_rate', hp.RealInterval(1e-4, 1e-2))
HP_BATCH_SIZE = hp.HParam('batch_size', hp.Discrete([64, 128]))
HP_EPOCH = hp.HParam('epoch', hp.Discrete([50,100]))
METRIC_ACCURACY = hp.Metric('accuracy')
# 定义超参数搜索空间
param_dist = {
    'lr': [0.001, 0.003,0.01 ,0.1],
    'batch_size': [64,128],
    'num_epochs': [50,100]
}


def train(epoch):
    model.train()
    

    for batch_idx, (data, target) in enumerate(train_loader):
        data, target = Variable(data), Variable(target)
        
        if torch.cuda.is_available():
            data = data.cuda()
            target = target.cuda()
        
        optimizer.zero_grad()
        output = model(data)
        loss = criterion(output, target)
        
        loss.backward()
        optimizer.step()
        
        if (batch_idx + 1)% 100 == 0:
            print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
                epoch, (batch_idx + 1) * len(data), len(train_loader.dataset),
                100. * (batch_idx + 1) / len(train_loader), 
                loss.item()))
    exp_lr_scheduler.step()
    writer.add_scalar('Loss/train', loss/len(train_loader), epoch)
    writer.flush()
def evaluate(data_loader):
    model.eval()
    loss = 0
    correct = 0
    
    for data, target in data_loader:
        data, target = Variable(data, volatile=True), Variable(target)
        if torch.cuda.is_available():
            data = data.cuda()
            target = target.cuda()
        
        output = model(data)
        
        loss += F.cross_entropy(output, target, size_average=False).item()

        pred = output.data.max(1, keepdim=True)[1]
        correct += pred.eq(target.data.view_as(pred)).cpu().sum()
        
    loss /= len(data_loader.dataset)
    writer.add_scalar('accuracy/train', correct / len(data_loader.dataset), epoch)
    writer.flush()    
    print('\nAverage loss: {:.4f}, Accuracy: {}/{} ({:.3f}%)\n'.format(
        loss, correct, len(data_loader.dataset),
        100. * correct / len(data_loader.dataset)))
    return correct / len(data_loader.dataset)

hparams_dir='logs/hparam_tuning'
hparams_writer = SummaryWriter(hparams_dir)
# 进行随机超参数搜索
param_list = list(ParameterSampler(param_dist, n_iter=20))
model = Net()
criterion = nn.CrossEntropyLoss()

for params in param_list:
    hparams_dict ={
        HP_LEARNING_RATE.name :params["lr"],
        HP_BATCH_SIZE.name :params['batch_size'],
        HP_EPOCH.name :params['num_epochs']

    }
    
    batch_size = params['batch_size']
    train_dataset = MNIST_data('../datasets/digit-recognizer/train.csv', n_pixels =n_pixels,transform= transforms.Compose(
                                [transforms.ToPILImage(), RandomRotation(degrees=20), RandomShift(3),
                                transforms.ToTensor(), transforms.Normalize(mean=(0.5,), std=(0.5,))]))
    train_loader = torch.utils.data.DataLoader(dataset=train_dataset,
                                            batch_size=batch_size, shuffle=True)

    logdir = "logs/" + 'epoch{0}_lr{1}_batch_size{2}.pth'.format(params['lr'],params['num_epochs'],params['batch_size'])
    writer = SummaryWriter(logdir)
    optimizer = optim.Adam(model.parameters(), lr=params['lr'])
    exp_lr_scheduler = lr_scheduler.StepLR(optimizer, step_size=5, gamma=0.5)

    if torch.cuda.is_available():
        model = model.cuda()
        criterion = criterion.cuda()

    n_epochs = params['num_epochs']

    for epoch in range(n_epochs):
        train(epoch)
        accuracy=evaluate(train_loader)
        if epoch==n_epochs-1:
            hparams_writer.add_hparams(hparams_dict,{'Accuracy':accuracy})
    torch.save(model.state_dict(), 'epoch{0}_lr{1}_batch_size{2}.pth'.format(params['num_epochs'],params['lr'],params['batch_size']))
    writer.close()
hparams_writer.close()

2远端电脑启动tensorboard

tensorboard --logdir logs

如果您的 TensorBoard 日志存储在远程服务器上，但您无法通过本地计算机上的浏览器访问它，可能是由于防火墙或网络设置的限制导致的。以下是一些可能的解决方案：

使用 SSH 隧道：如果您无法直接访问远程服务器上的 TensorBoard 日志，请考虑使用 SSH 隧道来建立本地和远程服务器之间的安全连接。在终端中使用以下命令建立 SSH 隧道：

   ssh -L 6006:localhost:6006 username@remote_server_ip

其中 username 是您在远程服务器上的用户名，remote_server_ip 是远程服务器的 IP 地址。然后，在本地计算机上打开浏览器并访问 http://localhost:6006 即可访问 TensorBoard 日志。

如果您使用的是 Windows 操作系统，可以使用 PuTTY 或其他 SSH 客户端来建立 SSH 隧道。
实例