训练的时候LOSS始终不会下降

Question

最近刚接触深度学习，想要自己做一下模型训练，使用NIN网络训练一个手写字符的数据集，写完后不知道为什么LOSS一直掉不下去，可以说是根本没有下降，NIN网络以及batchsize和lr都是按照李沐花书参数来的，死活找不到问题所在，恳请各位指条明路

训练代码如下

import torchvision
import torch.nn as nn
from torch.utils.data import DataLoader
from model1 import *

from torch.nn import CrossEntropyLoss
from torch.utils.tensorboard import SummaryWriter
from torch import argmax
transform = torchvision.transforms.Compose([torchvision.transforms.Resize((224,224)),
                                            torchvision.transforms.ToTensor()]
                                           )

train_dataset = torchvision.datasets.MNIST("MNI_datasets",train=True,transform=transform,download=True)
test_dataset = torchvision.datasets.MNIST("MNI_datasets",train=False,transform=transform,download=True)

train_dataload = DataLoader(train_dataset,batch_size=32)
test_dataload = DataLoader(test_dataset,batch_size=32)

test_len = len(test_dataset)
train_len = len(train_dataset)

writer = SummaryWriter("logs")
nin = Nin()
nin = nin.cuda()
loss_fn = CrossEntropyLoss()
loss_fn = loss_fn.cuda()
lr = 0.1
opti = torch.optim.SGD(nin.parameters(),lr = lr)

train_step = 0
test_step = 0
epoch = 10
#开始训练
nin.train()
for i in range(epoch):
    print("------第{}轮训练开始------".format(i+1))
    train_total_loss = 0.0
    test_total_loss = 0.0
    total_accuracy = 0.0
    for data in train_dataload:
        imgs,targets = data
        imgs = imgs.cuda()
        targets = targets.cuda()
        output = nin(imgs)
        loss = loss_fn(output,targets)
        opti.zero_grad()
        loss.backward()
        opti.step()
        train_step+=1
        train_total_loss = train_total_loss+loss
        
        if train_step%100==0:
            #writer.add_scalar("test_line",loss.item(),train_step)
            print("训练次数为{}，训练损失为{}".format(train_step,loss))
    #测试集
    with torch.no_grad():
        for data in test_dataload:
            imgs, targets = data
            imgs = imgs.cuda()
            targets = targets.cuda()
            output = nin(imgs)
            loss = loss_fn(output,targets)
            test_total_loss = loss.item()+test_total_loss
            accuracy = (output.argmax(1)==targets).sum()
            
            total_accuracy = accuracy+total_accuracy
            test_step+=1
            if test_step%100==0:
                #writer.add_scalar("test_line",loss.item(),test_step)
                print("测试次数为{},测试损失为{}".format(test_step,loss))
    print("本轮训练总损失为{}".format(train_total_loss))
    print("本轮测试总损失为{}".format(test_total_loss))
    print("本轮测试准确率为{}".format(total_accuracy/test_len))

    #torch.save(nin,"nin_{}.pth".format(i+1))

#writer.close()

  

网络模型如下

import torch.nn as nn
import torch
from torch.nn import Conv2d,ReLU,MaxPool2d,Dropout,AdaptiveMaxPool2d,Flatten

def Nin_block(in_channels,out_channels,kernel_size,stride,padding):
    return nn.Sequential(
        Conv2d(in_channels,out_channels,kernel_size,padding=padding,stride=stride),
        ReLU(),
        Conv2d(out_channels,out_channels,kernel_size=1),
        ReLU(),
        Conv2d(out_channels,out_channels,kernel_size=1),
        ReLU()
    )

class Nin(nn.Module):
    def __init__(self):
        super(Nin,self).__init__()
        self.model = nn.Sequential(
            Nin_block(1,96,11,4,0),
            MaxPool2d(3,2),
            Nin_block(96,256,5,1,2),
            MaxPool2d(3,2),
            Nin_block(256,384,3,1,1),
            MaxPool2d(3,2),
            Dropout(0.5),
            Nin_block(384,10,3,1,1),
            AdaptiveMaxPool2d((1,1)),
            Flatten()
        )

    def forward(self,input):
        output = self.model(input)
        return output

if __name__ =='__main__':
    nin = Nin()
   

Answer 1

可以适当增加层数，看着你这里只有一种，可以尝试添加全链接层