今天在自己手敲一遍线性回归代码的时候碰到了个bug。代码和李沐老师课上的相同：
```python
import numpy as np
import torch
import random
def generate_data(weight, bias, num_of_data):
data = torch.normal(1, 0.001, (num_of_data, weight.shape[0]))
result = torch.matmul(data, weight) + bias
label = result + torch.normal(1, 0.001, result.shape)
return data, label
def data_iter(batch_size, data, label):
number = len(data)
indice = list(range(number))
random.shuffle(indice)
for i in range(0, number, batch_size):
batch_indices = torch.tensor(indice[i:min(i+batch_size, number)])
yield data[batch_indices], label[batch_indices]
def linear_model(x, w, b):
return torch.matmul(x,w)+b`
def sgd(lr, batch_size, params):
with torch.no_grad():
for param in params:
param -= param.grad * lr / batch_size
param.grad.zero_()
def square_loss(predict, label):
return (predict-label)**2`
num_of_data = 100
weight = torch.normal(1, 0.001, (2,1), requires_grad=True)
bias = torch.zeros(1, requires_grad=True)
data, label = generate_data(weight, bias, num_of_data)
epoch = 10
batch_size = 10
learning_rate = 0.01
for i in range(epoch):
for X, Y in data_iter(batch_size, data, label):
output = linear_model(X, weight, bias)
l = square_loss(output, Y)
l.sum().backward()
sgd(learning_rate, batch_size, [weight, bias])
with torch.no_grad():
print(f'epoch: {epoch}, loss: {square_loss(output(data, weight, bias), label)}')`
```
在l.sum().backward()这一步遇到了问题：Trying to backward through the graph a second time, but the buffers have already been freed. Specify retain_graph=True when calling backward the first time. 尝试解决，但是确实没看出来哪里出现了错误。为什么会出现重复使用计算图的情况呢… 感谢各位！
