import torch import matplotlib.pyplot as plt #画图的 import random #产生随机数 def create_data(w, b, data_num): #生成数据 x = torch.normal(0, 1, (data_num, len(w))) #生成随机数x y = torch.matmul(x, w)+b #matmul表示矩阵相乘 noise = torch.normal(0, 0.01, y.shape) #噪声加在y上 y += noise return x, y num = 500 true_w = torch.tensor([8.1,2,2,4]) #必须用张量才能相乘 true_b = torch.tensor(1.1) X, Y = create_data(true_w, true_b, num) plt.scatter(X[:, 3], Y, 1) plt.show() def data_provider(data, label, batchsize): #每次访问这个函数，就能提供一批数据 length = len(label) indices = list(range(length)) random.shuffle(indices) #不能按顺序取数据，把数据打乱 for each in range(0,length, batchsize): get_indices = indices[each:each+batchsize] get_data = data[get_indices] get_label = label[get_indices] yield get_data, get_label #存档点的return 分批次返回数据，每次从这里取一个batch_size的数据，下次从这里继续取 batchsize = 16 # for batch_x, batch_y in data_provider(X, Y, batchsize): # print(batch_x, batch_y) # break def fun(x, w, b): pred_y = torch.matmul(x, w)+b return pred_y def maeLoss(pre_y, y): #计算loss：预测值和真实值 return torch.sum(abs(pre_y-y))/len(y) def sgd(paras, lr): #随机梯度下降，更新参数，梯度保存在参数中 with torch.no_grad(): #属于这句代码的部分，不计入梯度 for para in paras: para-=para.grad* lr para.grad.zero_() #使用过的梯度，归0 lr = 0.03 w_0 = torch.normal(0, 0.01, true_w.shape, requires_grad=True) #这个w需要计算梯度 b_0 = torch.tensor(0.01, requires_grad=True) print(w_0, b_0) epochs = 50 for epoch in range(epochs): data_loss = 0 for batch_x, batch_y in data_provider(X, Y, batchsize): #取一批数据 pred_y=fun(batch_x, w_0, b_0) loss = maeLoss(pred_y,batch_y) loss.backward() #梯度回传 sgd([w_0, b_0], lr) #更新参数 data_loss += loss #统计一轮所有loss，loss越小越好 print("epoch %03d: loss: %.6f"%(epoch, data_loss)) print("真实的函数值是" , true_w, true_b) print("训练得到的参数值是", w_0, b_0) idx = 2 #画图，可视化 plt.plot(X[:, idx].detach().numpy(), X[:, idx].detach().numpy()*w_0[idx].detach().numpy()+b_0.detach().numpy()) #只能取某一列作为自变量 plt.scatter(X[:, idx], Y, 1) plt.show()