[PyTorch practice] Linear regression of zero-based entry PyTorch [line-by-line code explanation]

foreword

This section contains a small example of getting started with torch

The line-by-line code video explanation is at: https://www.bilibili.com/video/BV1nS4y1u76S?spm_id_from=333.999.0.0

Mainly about linear regression examples

Create your own data
Build a Linear Regression Model
complete the training process
Drawing display

Linear model

where k is the weight and b is the bias term.

In general, the linear model is to fit the k and b of which are actually w and bias

like here

code as a whole

Simulation data

Add Gaussian white noise (a group of random numbers that conform to a normal distribution with a mean of 0 and a variance of 1), and set x to 512 points, that is, the number of samples is 512

Linear model

Because each value of the input and each value of the output is actually a dimension of 1, feature_num=1, the linear model is

class LinearModel(nn.Module):
    def __init__(self, in_fea, out_fea):
        super(LinearModel, self).__init__()
        self.out = nn.Linear(in_fea, out_fea)
    def forward(self, x):
        x = self.out(x)
        return x

Define loss function and optimizer

optimizer = torch.optim.SGD(model.parameters(), lr=0.02)

loss_func = nn.MSELoss()

Change the dimension of the data to the model input dimension

Add one dimension to the dimension of the feature

training and visualization

The routine is

forward reasoning
算loss
clear gradient
backpropagation
update weights

plt.ion()
for step in range(200):
    prediction = model(x)
    loss = loss_func(prediction, y)
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

    if step%10 == 0:
        plt.cla()
        plt.scatter(x.data.numpy(), y.data.numpy())
        plt.plot(x.data.numpy(), prediction.data.numpy(), 'r-', lw=5)
        plt.xlim(0,1.1)
        plt.ylim(0, 20)
        [w, b] = model.parameters()
        plt.text(0, 0.5, 'loss=%.4f, k=%.2f, b=%2f'%(loss.item(), w.item(), b.item() ),fontdict={'size': 20, 'color':  'red'})
        plt.pause(0.5)
plt.ioff()
plt.show()

Among them, ion is to open the interactive mode, and ioff is to close the interactive mode, and you can dynamically draw the picture to see the changes.

Dynamically look at the fitted line transformation