Pytorch神经网络—自定义数据集上实现

第一步、导入需要的包

import os
import scipy.io as sio
import numpy as np
import torch
import torch.nn as nn
import torch.backends.cudnn as cudnn
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader
from torchvision import transforms, utils
from torch.autograd import Variable

batchSize = 128 # batchsize的大小
niter = 10 # epoch的最大值

第二步、构建神经网络

设神经网络为如上图所示，输入层4个节点，两层隐含层各4个节点，输出层一个节点。每一层网络所做的都是线性变换，即y=W×X+b；代码实现如下：

class Neuralnetwork(nn.Module):
def __init__(self, in_dim, n_hidden_1, n_hidden_2, out_dim):
super(Neuralnetwork, self).__init__()
self.layer1 = nn.Linear(in_dim, n_hidden_1)
self.layer2 = nn.Linear(n_hidden_1, n_hidden_2)
self.layer3 = nn.Linear(n_hidden_2, out_dim)

def forward(self, x):
x = x.view(x.size(0), -1)
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
return x

model = Neuralnetwork(1*3, 4, 4, 1)
print(model) # net architecture

Neuralnetwork(
(layer1): Linear(in_features=3, out_features=4, bias=True)
(layer2): Linear(in_features=4, out_features=4, bias=True)
(layer3): Linear(in_features=4, out_features=1, bias=True)
)​​

第三步、读取数据

自定义的数据为demo_SBPFea.mat，是MATLAB保存的数据格式，其存储的内容如下：包括fea(1000*3)和sbp(1000*1)两个数组；fea为特征向量，行为样本数，列为特征宽度；sbp为标签

class SBPEstimateDataset(Dataset):

def __init__(self, ext='demo'):

data = sio.loadmat(ext+'_SBPFea.mat')
self.fea = data['fea']
self.sbp = data['sbp']

def __len__(self):

return len(self.sbp)
def __getitem__(self, idx):
fea = self.fea[idx]
sbp = self.sbp[idx]
"""Convert ndarrays to Tensors."""
return {'fea': torch.from_numpy(fea).float(),
'sbp': torch.from_numpy(sbp).float()
}

train_dataset = SBPEstimateDataset(ext='demo')
train_loader = DataLoader(train_dataset, batch_size=batchSize, # 分批次训练
shuffle=True, num_workers=int(8))

整个数据样本为1000，以batchSize = 128划分，分为8份，前7份为104个样本，第8份则为104个样本。在网络训练过程中，是一份数据一份数据进行训练的

第四步、模型训练

# 优化器，Adam
optimizer = optim.Adam(list(model.parameters()), lr=0.0001, betas=(0.9, 0.999),weight_decay=0.004)
scheduler = optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.997)
criterion = nn.MSELoss() # loss function
if torch.cuda.is_available(): # 有GPU，则用GPU计算
model.cuda()
criterion.cuda()

for epoch in range(niter):
losses = []
ERROR_Train = []
model.train()
for i, data in enumerate(train_loader, 0):
model.zero_grad()# 首先提取清零
real_cpu, label_cpu = data['fea'], data['sbp']

if torch.cuda.is_available():# CUDA可用情况下，将Tensor 在GPU上运行
real_cpu = real_cpu.cuda()
label_cpu = label_cpu.cuda()


input=real_cpu
label=label_cpu

inputv = Variable(input)
labelv = Variable(label)

output = model(inputv)
err = criterion(output, labelv)
err.backward()
optimizer.step()

losses.append(err.data[0])

error = output.data-label+ 1e-12
ERROR_Train.extend(error)

MAE = np.average(np.abs(np.array(ERROR_Train)))
ME = np.average(np.array(ERROR_Train))
STD = np.std(np.array(ERROR_Train))

print('[%d/%d] Loss: %.4f MAE: %.4f Mean Error: %.4f STD: %.4f' % (
epoch, niter, np.average(losses), MAE, ME, STD))​​

[0/10] Loss: 18384.6699 MAE: 135.3871 Mean Error: -135.3871 STD: 7.5580
[1/10] Loss: 17063.0215 MAE: 130.4145 Mean Error: -130.4145 STD: 7.8918
[2/10] Loss: 13689.1934 MAE: 116.6625 Mean Error: -116.6625 STD: 9.7946
[3/10] Loss: 8192.9053 MAE: 89.6611 Mean Error: -89.6611 STD: 12.9911
[4/10] Loss: 2979.1340 MAE: 52.5410 Mean Error: -52.5279 STD: 15.0930
[5/10] Loss: 599.7094 MAE: 22.2735 Mean Error: -19.9979 STD: 14.2069
[6/10] Loss: 207.2831 MAE: 11.2394 Mean Error: -4.8821 STD: 13.5528
[7/10] Loss: 189.8173 MAE: 9.8020 Mean Error: -1.2357 STD: 13.7095
[8/10] Loss: 188.3376 MAE: 9.6512 Mean Error: -0.6498 STD: 13.7075
[9/10] Loss: 186.8393 MAE: 9.6946 Mean Error: -1.0850 STD: 13.6332​