【深度学习】常见优化器的PyTorch实现
这里主要讲不同常见优化器代码的实现,以及在一个小数据集上做一个简单的比较。
备注:pytorch需要升级到最新版本
其中,SGD和SGDM,还有Adam是pytorch自带的优化器,而RAdam是最近提出的一个说是Adam更强的优化器,但是一般情况下真正的大佬还在用SGDM来做优化器。
导入必要库:
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import matplotlib.pyplot as plt
import torch.utils.data as Data
from torch.optim.optimizer import Optimizer
import math
主程序部分:
LR = 0.01
BATCH_SIZE = 32
EPOCH = 12
# fake dataset
x = torch.unsqueeze(torch.linspace(-1, 1, 300), dim=1)
y = x.pow(2) + 0.1 * torch.normal(torch.zeros(*x.size()))
torch_dataset = Data.TensorDataset(x, y)
loader = Data.DataLoader(
dataset=torch_dataset,
batch_size=BATCH_SIZE,
shuffle=True,
num_workers=2
)
class Net(nn.Module):
def __init__(self):
self).__init__()
nn.Linear(1, 20) =
nn.Linear(20, 1) =
def forward(self, x):
x = F.relu(self.hidden(x))
x = self.prediction(x)
return x
def main():
net_SGD = Net()
net_Momentum = Net()
net_Adam = Net()
net_RAdam = Net()
nets = [net_SGD, net_Momentum, net_Adam, net_RAdam]
opt_SGD = optim.SGD(net_SGD.parameters(), lr=LR)
opt_Momentum = optim.SGD(net_Momentum.parameters(), lr=LR, momentum=0.9)
opt_Adam = optim.Adam(net_Adam.parameters(), lr=LR, betas=(0.9, 0.99))
opt_RAdam = RAdam(net_RAdam.parameters(),lr=LR,weight_decay=0)
optimizers = [opt_SGD, opt_Momentum, opt_Adam, opt_RAdam]
loss_func = nn.MSELoss()
losses_his = [[], [], [], []]
# training
for epoch in range(EPOCH):
', epoch) :
for step, (batch_x, batch_y) in enumerate(loader):
b_x = batch_x
b_y = batch_y
for net, opt, l_his in zip(nets, optimizers, losses_his):
out = net(b_x)
loss = loss_func(out, b_y)
opt.zero_grad()
loss.backward()
opt.step()
l_his.append(loss.item())
labels = ['SGD', 'Momentum', 'Adam','RAdam']
for i, l_his in enumerate(losses_his):
label=labels[i])
'best') =
plt.xlabel('Steps')
plt.ylabel('Loss')
0.2))
plt.show()
if __name__ == '__main__':
main()
下图是优化器的对比:
可以看出来,Adam的效果可以说是非常好的。然后SGDM其次,SGDM是大佬们经常会使用的,所以在这里虽然看起来SGDM效果不如Adam,但是依然推荐在项目中,尝试一下SGDM的效果。
往期精彩回顾
获取一折本站知识星球优惠券,复制链接直接打开:
https://t.zsxq.com/yFQV7am
本站qq群1003271085。
加入微信群请扫码进群:
评论