python-pytorch基础之神经网络回归

定义数据集

import torch
import random

定义函数

# 生成数据
def get_rancledata():width=random.random()height=random.random()s=width*heightreturn width,height,sget_rancledata()

(0.1571327616035657, 0.5335562021159256, 0.08383915950918565)

生成数据集

class dataset(torch.utils.data.Dataset):def __init__(self):passdef __len__(self):return 1000def __getitem__(self,i):width,height,s=get_rancledata()x=torch.FloatTensor([width,height])# 这里注意也是需要转换成tensor的，否则训练会报类型错误y=torch.FloatTensor([s])return x,ydataset=dataset()

len(dataset),dataset[4999]

(1000, (tensor([0.2137, 0.6781]), tensor([0.1449])))

使用Dataloader加载dataset

loader=torch.utils.data.DataLoader(dataset=dataset,shuffle=True,batch_size=9
)

len(loader),next(iter(loader))

(112,[tensor([[0.7389, 0.1202],[0.5764, 0.7888],[0.7244, 0.0229],[0.5102, 0.0755],[0.8550, 0.4998],[0.9992, 0.5890],[0.1704, 0.0162],[0.2132, 0.9157],[0.7946, 0.8907]]),tensor([[0.0888],[0.4546],[0.0166],[0.0385],[0.4273],[0.5885],[0.0028],[0.1953],[0.7077]])])

定义神经网络

定义

# 输入两个，输出一个
class Model(torch.nn.Module):def __init__(self):super().__init__()self.fb=torch.nn.Sequential(torch.nn.Linear(in_features=2,out_features=32),torch.nn.ReLU(),torch.nn.Linear(in_features=32,out_features=32),torch.nn.ReLU(),torch.nn.Linear(in_features=32,out_features=1))def forward(self,x):return self.fb(x)

实例化

model=Model()

torch.rand(4,2)

tensor([[0.4510, 0.1455],[0.4963, 0.2974],[0.9480, 0.9913],[0.9053, 0.4228]])

查看是否是输出的一个

# 测试
model(torch.rand(8,2)).shape

torch.Size([8, 1])

训练

编写trian方法

def train():# 选择损失函数loss_fn=torch.nn.MSELoss()# 选择优化器optimizer=torch.optim.Adam(model.parameters(),lr=1e-4)#遍历多少轮for epoch in range(100):#全量遍历for i ,(x,y) in enumerate(loader):#计算损失#计算梯度#优化参数#优化梯度清零out=model(x)loss=loss_fn(out,y)loss.backward()optimizer.step()optimizer.zero_grad()if epoch % 20 ==0:print(epoch,loss.item())torch.save(model,"huigui.model")

训练并保存模型

train()

0 0.03260539472103119
0 0.06368591636419296
0 0.08260147273540497
0 0.04632813110947609
0 0.08333451300859451
0 0.10992465913295746
0 0.12929300963878632
0 0.061169371008872986
0 0.08229123800992966
0 0.0604255348443985
0 0.11475709825754166
0 0.13913851976394653
0 0.09228374809026718
0 0.10618235915899277
0 0.12170673906803131
0 0.05438697338104248
0 0.11730150133371353
0 0.07718850672245026
0 0.11877405643463135
0 0.0647420659661293
0 0.1062769666314125
0 0.08034960925579071
0 0.06462960690259933
0 0.029708124697208405
0 0.19415663182735443
0 0.022178875282406807
0 0.023824863135814667
0 0.06074700132012367
0 0.014404748566448689
0 0.015829702839255333
0 0.07006165385246277
0 0.0908271074295044
0 0.023783870041370392
0 0.09584006667137146
0 0.16521167755126953
0 0.09473344683647156
0 0.12153694033622742
0 0.030839459970593452
0 0.019292233511805534
40 8.071886259131134e-05
40 2.137169212801382e-05
40 0.00010651862248778343
40 7.332033419515938e-05
40 0.00010564295371295884
40 4.790672755916603e-05
40 3.7615245673805475e-05
40 3.413142985664308e-05
40 6.713613402098417e-05
40 0.0006545005016960204

测试模型结果

构造数据

# 从loader加载一批数据来测试x,y=next(iter(loader))
x,y

测试

# 方法一
# out=model(x)# 方法二 加载模型
model1=torch.load("huigui.model")
out=model1(x)# 打印在一起，便于观察,
# 这个cat函数很有用注意
torch.cat([out,y],dim=1)

结论

从上面结果看

    [ 0.6257,  0.6214],[ 0.5435,  0.5454],[ 0.0227,  0.0203],[-0.0044,  0.0033],[ 0.5257,  0.5296],[ 0.4749,  0.4805],[ 0.4665,  0.4649],[ 0.4143,  0.4141],[ 0.0130,  0.0138]]

第一列是预测的，第二列是实际的，可以查看两列值相差很小，说明模型有效