python - 为什么这两行代码在 Pytorch 中给了我不同的输出，这是否解释了奇怪的参数？

Question

我正在使用 Pytorch 运行具有两个输入和两个输出的线性回归。我生成了一个数据集，然后添加了噪声，这样我就可以练习计算权重和偏差。这都是通过我正在参加的在线课程完成的。

模型生成的权重与我初始化模型的权重完全不同，但是偏差是相同的。这令人困惑，因为损失基本上为零 (10^-7)。

我在有问题的代码周围加上问号。生成的输出如下。如您所见，训练模型找到 'Linear.weight': tensor([[1.3696, 0.6309], [1.3159, 0.6847]]), 'Linear.bias': tensor([ 1.0000, -1.0000]) 的参数与我创建数据集时使用的参数非常不同。'Linear.weight': tensor([[1.0, -1.0], [1.0, 3.0]]), 'Linear.bias': tensor([ 1.0, -1.0])

如果我运行下面的行，我会得到我试图预测的 y 值

print(data_set.y)

如果我运行下面的行，我会得到模型预测的 y 值，它与上面的输出相匹配。

print(model(data_set.x))

我的想法是下面的行将与上面的行相同，但是从输出来看，它们非常不同。我真的很困惑

print(torch.mm(data_set.x, model.state_dict()['Linear.weight']) + model.state_dict()['Linear.bias'])

import torch
import numpy as np
import matplotlib.pyplot as plt
from torch import nn,optim
from torch.utils.data import Dataset, DataLoader
torch.manual_seed(1)

# Generating the dataset
class Data(Dataset):
    def __init__(self):
            self.x=torch.zeros(20,2)
            self.x[:,0]=torch.arange(-1,1,0.1)
            self.x[:,1]=torch.arange(-1,1,0.1)
            self.w=torch.tensor([ [1.0,-1.0],[1.0,3.0]])
            self.b=torch.tensor([[1.0,-1.0]])
            self.f=torch.mm(self.x,self.w)+self.b

            self.y=self.f+0.001*torch.randn((self.x.shape[0],1))
            self.len=self.x.shape[0]

    def __getitem__(self,index):

        return self.x[index],self.y[index]

    def __len__(self):
        return self.len

# creating a dataset object
data_set=Data()

# creating a linear_regression class
class linear_regression(nn.Module):
    def __init__(self,input_size,output_size):
        super(linear_regression,self).__init__()
        self.Linear=nn.Linear(input_size,output_size)
    def forward(self,x):
        yhat=self.Linear(x)
        return yhat

# creating a model object, an optimizer object, loss function, and dataloder
model=linear_regression(2,2)
optimizer = optim.SGD(model.parameters(), lr = 0.1)
criterion = nn.MSELoss()
train_loader=DataLoader(dataset=data_set,batch_size=5)

# List for storing loss values and creating epochs variable
LOSS=[]
epochs=100

# for loop that trains the model
for epoch in range(epochs):
    for x,y in train_loader:
        #make a prediction
        yhat=model(x)
        #calculate the loss
        loss=criterion(yhat,y)
        #store loss/cost
        LOSS.append(loss.item())
        #clear gradient
        optimizer.zero_grad()
        #Backward pass: compute gradient of the loss with respect to all the learnable parameters
        loss.backward()
        #the step function on an Optimizer makes an update to its parameters
        optimizer.step()

# ?????????????????????????????????????
print("model parameters: ", model.state_dict())
print(data_set.y)
print(model(data_set.x))
print(torch.mm(data_set.x, model.state_dict()['Linear.weight']) + model.state_dict()['Linear.bias'])
# ?????????????????????????????????????

plt.plot(LOSS)
plt.xlabel("iterations ")
plt.ylabel("Cost/total loss ")
plt.show()

输出：

model parameters:  OrderedDict([('Linear.weight', tensor([[1.3696, 0.6309],
        [1.3159, 0.6847]])), ('Linear.bias', tensor([ 1.0000, -1.0000]))])
tensor([[-1.0015e+00, -3.0015e+00],
        [-8.0075e-01, -2.8008e+00],
        [-6.0065e-01, -2.6007e+00],
        [-4.0161e-01, -2.4016e+00],
        [-1.9913e-01, -2.1991e+00],
        [ 2.4440e-04, -1.9998e+00],
        [ 1.9934e-01, -1.8007e+00],
        [ 4.0081e-01, -1.5992e+00],
        [ 6.0044e-01, -1.3996e+00],
        [ 8.0117e-01, -1.1988e+00],
        [ 1.0018e+00, -9.9823e-01],
        [ 1.1999e+00, -8.0010e-01],
        [ 1.4001e+00, -5.9994e-01],
        [ 1.5994e+00, -4.0062e-01],
        [ 1.7992e+00, -2.0080e-01],
        [ 1.9999e+00, -1.3162e-04],
        [ 2.1992e+00,  1.9920e-01],
        [ 2.4003e+00,  4.0034e-01],
        [ 2.6003e+00,  6.0028e-01],
        [ 2.8017e+00,  8.0172e-01]])
tensor([[-1.0005e+00, -3.0005e+00],
        [-8.0048e-01, -2.8005e+00],
        [-6.0043e-01, -2.6004e+00],
        [-4.0037e-01, -2.4004e+00],
        [-2.0031e-01, -2.2003e+00],
        [-2.5582e-04, -2.0003e+00],
        [ 1.9980e-01, -1.8002e+00],
        [ 3.9986e-01, -1.6001e+00],
        [ 5.9991e-01, -1.4001e+00],
        [ 7.9997e-01, -1.2000e+00],
        [ 1.0000e+00, -9.9997e-01],
        [ 1.2001e+00, -7.9991e-01],
        [ 1.4001e+00, -5.9986e-01],
        [ 1.6002e+00, -3.9980e-01],
        [ 1.8003e+00, -1.9974e-01],
        [ 2.0003e+00,  3.1275e-04],
        [ 2.2004e+00,  2.0037e-01],
        [ 2.4004e+00,  4.0043e-01],
        [ 2.6005e+00,  6.0048e-01],
        [ 2.8005e+00,  8.0054e-01]], grad_fn=<AddmmBackward>)
tensor([[-1.6855, -2.3156],
        [-1.4169, -2.1840],
        [-1.1484, -2.0525],
        [-0.8798, -1.9209],
        [-0.6113, -1.7893],
        [-0.3427, -1.6578],
        [-0.0742, -1.5262],
        [ 0.1944, -1.3947],
        [ 0.4629, -1.2631],
        [ 0.7315, -1.1315],
        [ 1.0000, -1.0000],
        [ 1.2686, -0.8684],
        [ 1.5371, -0.7368],
        [ 1.8057, -0.6053],
        [ 2.0742, -0.4737],
        [ 2.3428, -0.3422],
        [ 2.6113, -0.2106],
        [ 2.8799, -0.0790],
        [ 3.1484,  0.0525],
        [ 3.4170,  0.1841]])

标签： pythonmachine-learningdeep-learningpytorchlinear-regression