PyTorch 实现批训练和 Optimizer 优化器

批训练

import torch
import torch.utils.data as DataBATCH_SIZE = 5x = torch.linspace(1, 10, 10)   # this is x data
y = torch.linspace(10, 1, 10)   # this is y datatorch_dataset = Data.TensorDataset(x, y)
loader = Data.DataLoader(dataset=torch_dataset,batch_size=BATCH_SIZE,shuffle=True,   # 是否随机打乱数据num_workers=2,  # 使用几个线程/进程
)if __name__ == "__main__":for epoch in range(3):for step, (batch_x, batch_y) in enumerate(loader):# training...print('Epoch: ', epoch, '| Step:', step, '| batch x:',batch_x.numpy(), '| batch y:', batch_y.numpy())

Optimizer 优化器

SGD

喝醉的人，回家的时候，摇摇晃晃，曲折无比，走了很多弯路

w + = − L e a r n i n g w += -Learning w+=−Learning \; r a t e ∗ d x rate * dx rate∗dx

Momentum

将这个喝醉的人放到斜坡上，这个人就会不自觉的依靠惯性向下走，走得弯路也就少了

m = b 1 ∗ m − L e a r n i n g m = b1 * m - Learning m=b1∗m−Learning \; r a t e ∗ d x rate * dx rate∗dx
W + = m W += m W+=m

AdaGrad

学习率上动了手脚，使得每个参数的更新都会有自己与众不同的学习效率。
给他一双不好走的鞋子，鞋子变成了向前走的阻力，逼着他向前直着走。

v + = d x 2 v += dx^2 v+=dx2
W + = − L e a r n i n g W += -Learning W+=−Learning \; r a t e ∗ d x / v rate * dx / \sqrt v rate∗dx/v

RMsProp

将 Momentum 与 AdaGrad 结合
m = b 1 ∗ m + ( 1 − b 1 ) ∗ d x m = b1 * m +(1-b1) * dx m=b1∗m+(1−b1)∗dx ------>Momentum
b 2 ∗ v + ( 1 − b 2 ) ∗ d x 2 b2 *v +(1-b2)*dx^2 b2∗v+(1−b2)∗dx2 ------------->AdaGrad
W + = − L e a r n i n g W += -Learning W+=−Learning \; r a t e ∗ m / v rate * m/ \sqrt v rate∗m/v

import torch
import torch.utils.data as Data
import torch.nn.functional as F
from torch.autograd import Variable
import matplotlib.pyplot as plt# hyper parameters
LR = 0.01
BATCH_SIZE = 32
EPOCH = 12x = torch.unsqueeze(torch.linspace(-1, 1, 1000), dim=1)
y = x.pow(2) + 0.1 * torch.normal(torch.zeros(*x.size()))# plot dataset
# plot.scatter(x.numpy(), y.numpy())
# plt.show()torch_dataset = Data.TensorDataset(x, y)
loader = Data.DataLoader(dataset=torch_dataset, batch_size=BATCH_SIZE, shuffle=True)# default network
class Net(torch.nn.Module):def __init__(self):super(Net, self).__init__()self.hidden = torch.nn.Linear(1, 20)  # hidden layerself.predict = torch.nn.Linear(20, 1)  # output layerdef forward(self, x):x = F.relu(self.hidden(x))  # activation function for hidden layerx = self.predict(x)  # linear outputreturn x# different nets
net_SGD = Net()
net_Momentum = Net()
net_RMSprop = Net()
net_Adam = Net()
nets = [net_SGD, net_Momentum, net_RMSprop, net_Adam]opt_SGD = torch.optim.SGD(net_SGD.parameters(), lr=LR)
opt_Momentum = torch.optim.SGD(net_Momentum.parameters(), lr=LR, momentum=0.8)
opt_RMSprop = torch.optim.RMSprop(net_RMSprop.parameters(), lr=LR, alpha=0.9)
opt_Adam = torch.optim.Adam(net_Adam.parameters(), lr=LR, betas=(0.9, 0.99))
optimizers = [opt_SGD, opt_Momentum, opt_RMSprop, opt_Adam]loss_func = torch.nn.MSELoss()
losses_his = [[], [], [], []]  # record lossfor epoch in range(EPOCH):print(epoch)for step, (batch_x, batch_y) in enumerate(loader):b_x, b_y = Variable(batch_x), Variable(batch_y)for net, opt, l_his in zip(nets, optimizers, losses_his):output = net(b_x)  # get output for every netloss = loss_func(output, b_y)  # compute loss for every netopt.zero_grad()  # clear gradients for next trainloss.backward()  # backpropagation, compute gradientsopt.step()  # apply gradientsl_his.append(loss.item())  # loss recoderlabels = ['SGD', 'Momentum', 'RMSprop', 'Adam']
for i, l_his in enumerate(losses_his):plt.plot(l_his, label=labels[i])
plt.legend(loc='best')
plt.xlabel('Steps')
plt.ylabel('Loss')
plt.ylim((0, 0.2))
plt.show()