MINIST深度学习识别：python全连接神经网络和pytorch LeNet CNN网络训练实现及比较（二）

版权声明：本文为博主原创文章，欢迎转载，并请注明出处。联系方式：460356155@qq.com

在前一篇文章MINIST深度学习识别：python全连接神经网络和pytorch LeNet CNN网络训练实现及比较（一）中，直接用python创建全连接神经网络模型进行深度学习训练，这样可以对神经网络有较为深刻的认识。

但是在实际应用中，一般都是采用各种深度学习框架来开展人工智能项目，以下就采用pytorch来实现前一篇文章中的全连接神经网络（784-300-10）。

# -*- coding:utf-8 -*-

u"""pytorch LineNet神经网络训练学习MINIST"""

__author__ = 'zhengbiqing 460356155@qq.com'


import torch as t
import torchvision as tv
import torch.nn as nn
import torch.nn.functional as F
import torchvision.transforms as transforms
from torch.autograd import Variable
import matplotlib.pyplot as plt
import datetime


#是否训练网络
TRAIN = True

#是否保存网络
SAVE_PARA = False

#学习率和训练次数
LR = 0.05
EPOCH = 10

#训练每批次的样本数
BATCH_SZ = 16

#样本读取线程数
WORKERS = 4

#网络参赛保存文件名
PARAS_FN = 'minist_linenet_params.pkl'

#minist数据存放位置
ROOT = '/home/zbq/pytorch/minist'


#定义网络模型
class LineNet(nn.Module):
    def __init__(self):
        super(LineNet, self).__init__()

        self.fc = nn.Sequential(
            nn.Linear(28*28, 300),
            nn.ReLU(),
            nn.Linear(300, 10)
        )

    def forward(self, x):
        #x是2维tensor，转换为1维向量
        x = x.view(x.size()[0], -1)
        x = self.fc(x)
        return x


'''
训练并测试网络
net：网络模型
train_data_load：训练数据集
test_data_load：测试数据集
epochs：训练迭代次数
save：是否保存训练结果
'''
def net_train(net, train_data_load, test_data_load, epochs, save):
    start_time = datetime.datetime.now()

    loss_list = []

    for epoch in range(epochs):
        for i, data in enumerate(train_data_load, 0):
            img, label = data
            img, label = Variable(img), Variable(label)
            img, label = img.cuda(), label.cuda()

            optimizer.zero_grad()

            pre = net(img)
            loss = loss_func(pre, label)
            loss.backward()

            optimizer.step()

            #显示损失函数值的变化
            loss_data = loss.data.item()
            if i % 1000 == 999:
                print('epoch:{epoch} i:{i} loss:{loss}'.format(epoch=epoch, i=i, loss=loss_data))

            if i % 100 == 99:
                loss_list.append(loss_data)

        # 每个epoch结束后用测试集检查识别准确度
        net_test(epoch, test_data_load)

    print('MINIST pytorch LineNet Train: EPOCH:{epochs}, BATCH_SZ:{batch_sz}, LR:{lr}'.format(epochs=epochs, batch_sz=BATCH_SZ, lr=LR))
    print('train spend time: ', datetime.datetime.now() - start_time)

    if save:
        t.save(net.state_dict(), PARAS_FN)

    #显示目标函数值的变化曲线
    plt.plot(loss_list)
    plt.show()


'''
用测试集检查准确率
'''
def net_test(epoch, test_data_load):
    ok = 0

    for i, data in enumerate(test_data_load):
        img, label = data
        img, label = Variable(img), Variable(label)
        img, label = img.cuda(), label.cuda()

        outs = net(img)
        _, pre = t.max(outs.data, 1)
        ok += (pre == label).sum()

    acc = ok.item() * 100 / (len(test_data_load) * BATCH_SZ)

    print('EPOCH:{epoch}, ACC:{acc}\n'.format(epoch=epoch, acc=acc))


#图像数值转换，ToTensor源码注释
"""Convert a ``PIL Image`` or ``numpy.ndarray`` to tensor.
    Converts a PIL Image or numpy.ndarray (H x W x C) in the range
    [0, 255] to a torch.FloatTensor of shape (C x H x W) in the range [0.0, 1.0].
    """
#归一化，把[0.0, 1.0]变换为[-1,1], ([0, 1] - 0.5) / 0.5 = [-1, 1]
transform = tv.transforms.Compose([transforms.ToTensor(), transforms.Normalize([0.5], [0.5])])

#定义数据集
train_data = tv.datasets.MNIST(root=ROOT, train=True, download=True, transform=transform)
test_data = tv.datasets.MNIST(root=ROOT, train=False, download=False, transform=transform)

train_load = t.utils.data.DataLoader(train_data, batch_size=BATCH_SZ, shuffle=True, num_workers=WORKERS)
test_load = t.utils.data.DataLoader(test_data, batch_size=BATCH_SZ, shuffle=False, num_workers=WORKERS)

print('train data num:', len(train_data), ', test data num:', len(test_data))


net = LineNet()
net.cuda()

loss_func = nn.CrossEntropyLoss()
optimizer = t.optim.SGD(net.parameters(), lr=LR)

if TRAIN:
    net_train(net, train_load, test_load, EPOCH, SAVE_PARA)
else:
    net.load_state_dict(t.load(PARAS_FN))
    net_test(0, test_load)

网络训练结果准确率基本在97%～98%，和前一篇MINIST深度学习识别：python全连接神经网络和pytorch LeNet CNN网络训练实现及比较（一）相同网络结构的全连接神经网络相当，但是因为这里采用GPU运算，训练时间降低到1/8。

此外，借助pytorch，代码更简单。

运行结果如下：

train data num: 60000 , test data num: 10000
epoch:0 i:999 loss:0.3457891643047333
epoch:0 i:1999 loss:0.09639787673950195
epoch:0 i:2999 loss:0.27898865938186646
EPOCH:0, ACC:94.84

epoch:1 i:999 loss:0.33745211362838745
epoch:1 i:1999 loss:0.11106520891189575
epoch:1 i:2999 loss:0.21725007891654968
EPOCH:1, ACC:96.42

epoch:2 i:999 loss:0.3825737535953522
epoch:2 i:1999 loss:0.02866300940513611
epoch:2 i:2999 loss:0.11832481622695923
EPOCH:2, ACC:96.77

epoch:3 i:999 loss:0.11886310577392578
epoch:3 i:1999 loss:0.012149035930633545
epoch:3 i:2999 loss:0.030409961938858032
EPOCH:3, ACC:97.2

epoch:4 i:999 loss:0.008915185928344727
epoch:4 i:1999 loss:0.008089780807495117
epoch:4 i:2999 loss:0.0005310177803039551
EPOCH:4, ACC:97.6

epoch:5 i:999 loss:0.02993696928024292
epoch:5 i:1999 loss:0.01784616708755493
epoch:5 i:2999 loss:0.10544028878211975
EPOCH:5, ACC:97.6

epoch:6 i:999 loss:0.008486062288284302
epoch:6 i:1999 loss:0.0334945023059845
epoch:6 i:2999 loss:0.00291365385055542
EPOCH:6, ACC:97.37

epoch:7 i:999 loss:0.0062919557094573975
epoch:7 i:1999 loss:0.0003241896629333496
epoch:7 i:2999 loss:0.0006818175315856934
EPOCH:7, ACC:97.23

epoch:8 i:999 loss:0.0007421970367431641
epoch:8 i:1999 loss:0.005641639232635498
epoch:8 i:2999 loss:0.005949795246124268
EPOCH:8, ACC:97.7

epoch:9 i:999 loss:0.024028539657592773
epoch:9 i:1999 loss:0.005388796329498291
epoch:9 i:2999 loss:0.0029097795486450195
EPOCH:9, ACC:97.39

MINIST pytorch LineNet Train: EPOCH:10, BATCH_SZ:16, LR:0.05
train spend time:  0:00:43.183836

损失函数值变化曲线为：