关于#pytorch#的问题：pytorch实现mnist手写数字识别

Question

我正在用pytorch实现mnist手写数字识别，但是我的loss从一开始就不变，这是为什么？ 以下是我的代码：

import gzip
import pickle
import matplotlib.pyplot as plt
import numpy as np
import torch
from torch import nn
from torch import optim
from torch.nn import functional as F
from torch.autograd import variable

#读取训练数据
f = gzip.open("./mnist.pkl.gz","rb")
train_data, val_data, test_data = pickle.load(f,encoding='latin1')
f.close()

# 将50000张训练图片分为250组，每组200张图片，图片大小
train_data_img = train_data[0].reshape(250,200,28,28)
train_data_ans = train_data[1].reshape(250,200)
#搭建网络
class Net(nn.Module):
    def __init__(self):
        super(Net,self).__init__()
        self.fc1 = nn.Linear(784, 64)
        self.fc2 = nn.Linear(64, 32)
        self.fc3 = nn.Linear(32, 10)

    def forward(self,x):
        x = F.leaky_relu(self.fc1(x))
        x = F.leaky_relu(self.fc2(x))
        x = self.fc3(x)
        return x

net = Net()
optimizer = optim.SGD(net.parameters(),lr=0.01,momentum=0.0)

#迭代
losses = []
loss = 1
for epoch in range(10):
    for data,ans in zip(train_data_img,train_data_ans):
        out = net(torch.tensor(data.reshape(200,784)))
        ans = F.one_hot(torch.tensor(ans))
        loss = F.mse_loss(out, ans)
        optimizer.zero_grad()
        loss.backword()
        optimizer.step()
    losses.append(loss.item())

print(losses)
xlabel = np.linspace(0,len(losses),len(losses))
plt.plot(xlabel,losses)
plt.show()

Answer 1

可以将每一次的w、b、loss、dw、db都打印出来，看是否随机梯度下降没起作用