[PyTorch小试牛刀]实战五·RNN(LSTM)实现逻辑回归对FashionMNIST数据集进行分类（使用GPU）

[PyTorch小试牛刀]实战五·RNN(LSTM)实现逻辑回归对FashionMNIST数据集进行分类（使用GPU）

内容还包括了网络模型参数的保存于加载。
数据集
下载地址
代码部分

import torch as t
import torchvision as tv
import numpy as np
import time


# 超参数
EPOCH = 5
BATCH_SIZE = 100
DOWNLOAD_MNIST = True   # 下过数据的话, 就可以设置成 False
N_TEST_IMG = 10          # 到时候显示 

TIME_STEP = 28      # rnn 时间步数 / 图片高度
INPUT_SIZE = 28     # rnn 每步输入值 / 图片每行像素


class NN(t.nn.Module):
    def __init__(self):
        super(NN, self).__init__()

        train_data = tv.datasets.FashionMNIST(
        root="./fashionmnist/",
        train=True,
        transform=tv.transforms.ToTensor(),
        download=DOWNLOAD_MNIST
        )

        test_data = tv.datasets.FashionMNIST(
        root="./fashionmnist/",
        train=False,
        transform=tv.transforms.ToTensor(),
        download=DOWNLOAD_MNIST
        )

        #print(test_data)


        # Data Loader for easy mini-batch return in training, the image batch shape will be (50, 1, 28, 28)
        self.train_loader = t.utils.data.DataLoader(
            dataset=train_data, 
            batch_size=BATCH_SIZE,
            shuffle=True)

        self.test_loader = t.utils.data.DataLoader(
            dataset=test_data, 
            batch_size=1000,
            shuffle=True)

        self.rnn = t.nn.Sequential(
            t.nn.LSTM(     # LSTM 效果要比 nn.RNN() 好多了
            input_size=28,      # 图片每行的数据像素点
            hidden_size=256,     # rnn hidden unit
            num_layers=2,       # 有几层 RNN layers
            batch_first=True,   # input & output 会是以 batch size 为第一维度的特征集 e.g. (batch, time_step, input_size)
            )                   # output shape (16, 28, 28)
        )

        self.dnn = t.nn.Linear(256,10)

        self.lr = 0.001
        self.loss = t.nn.CrossEntropyLoss()
        self.opt = t.optim.Adam(self.parameters(), lr = self.lr)

    def forward(self,x):
        # x shape (batch, time_step, input_size)
        # r_out shape (batch, time_step, output_size)
        # h_n shape (n_layers, batch, hidden_size)   LSTM 有两个 hidden states, h_n 是分线, h_c 是主线
        # h_c shape (n_layers, batch, hidden_size)

        rnn1 = self.rnn(x)
        #print(cnn1.shape)
        r_out, (h_n, h_c)  = rnn1
        #print(cnn1.shape)
        out = self.dnn(r_out[:,-1,:])

        return(out)

def train():
    use_gpu =   t.cuda.is_available()
    model = NN()
    if(use_gpu):
        model.cuda()
    print(model)
    loss = model.loss
    opt = model.opt
    dataloader = model.train_loader
    testloader = model.test_loader

    
    for e in range(EPOCH):
        step = 0
        ts = time.time()
        for (x, y) in (dataloader):
            

            model.train()# train model dropout used
            step += 1
            b_x = x.view(-1,28,28)   # batch x, shape (batch, 28*28)
            #print(b_x.shape)
            b_y = y
            if(use_gpu):
                b_x = b_x.cuda()
                b_y = b_y.cuda()
            out = model(b_x)
            losses = loss(out,b_y)
            opt.zero_grad()
            losses.backward()
            opt.step()
            if(step%100 == 0):
                if(use_gpu):
                    print(e,step,losses.data.cpu().numpy())
                else:
                    print(e,step,losses.data.numpy())
                
                model.eval() # train model dropout not use
                for (tx,ty) in testloader:
                    t_x = tx.view(-1,28,28)   # batch x, shape (batch, 28*28)
                    t_y = ty
                    if(use_gpu):
                        t_x = t_x.cuda()
                        t_y = t_y.cuda()
                    t_out = model(t_x)
                    if(use_gpu):
                        acc = (np.argmax(t_out.data.cpu().numpy(),axis=1) == t_y.data.cpu().numpy())
                    else:
                        acc = (np.argmax(t_out.data.numpy(),axis=1) == t_y.data.numpy())

                    print(time.time() - ts ,np.sum(acc)/1000)
                    ts = time.time()
                    break#只测试前1000个
            


    t.save(model, './model.pkl')  # 保存整个网络
    t.save(model.state_dict(), './model_params.pkl')   # 只保存网络中的参数 (速度快, 占内存少)
    #加载参数的方式
    """net = DNN()
    net.load_state_dict(t.load('./model_params.pkl'))
    net.eval()"""
    #加载整个模型的方式
    net = t.load('./model.pkl')
    net.cpu()
    net.eval()
    for (tx,ty) in testloader:
        t_x = tx.view(-1,28,28)   # batch x, shape (batch, 28*28)
        t_y = ty

        t_out = net(t_x)
        #acc = (np.argmax(t_out.data.CPU().numpy(),axis=1) == t_y.data.CPU().numpy())
        acc = (np.argmax(t_out.data.numpy(),axis=1) == t_y.data.numpy())

        print(np.sum(acc)/1000)

if __name__ == "__main__":
    train()

输出结果

NN(
  (rnn): Sequential(
    (0): LSTM(28, 256, num_layers=2, batch_first=True)
  )
  (dnn): Linear(in_features=256, out_features=10, bias=True)
  (loss): CrossEntropyLoss()
)
0 100 0.77180815
3.650240659713745 0.706
0 200 0.8147288
3.3065454959869385 0.711
0 300 0.754965
3.3209893703460693 0.736
0 400 0.5886362
3.3486075401306152 0.803
0 500 0.4883507
3.3163959980010986 0.781
0 600 0.66265166
3.3470709323883057 0.808
1 100 0.3800248
3.3159289360046387 0.821
1 200 0.30893803
3.403984785079956 0.826
1 300 0.59795433
3.7441184520721436 0.84
1 400 0.48738843
3.3226170539855957 0.854
1 500 0.392042
3.3506269454956055 0.843
1 600 0.25022513
。。。
3.291714906692505 0.871
4 500 0.3532069
3.344895839691162 0.88
4 600 0.2680706
3.7954905033111572 0.882
0.888
0.886
0.89
0.859
0.874
0.881
0.869
0.888
0.866
0.885

结果分析
我笔记本配置为CPU i5 8250u GPU MX150 2G内存
使用CPU训练时，每100步，58秒左右
使用GPU训练时，每100步，3.3秒左右
提升了将近20倍，
经过测试，使用GPU运算RNN速率大概是CPU的15~20倍，推荐大家使用GPU运算，就算GPU配置差些也可以显著提升效率。