Optimizer · Apache SINGA
5.0.0_Chinese
Docs
Community
News
API
GitHub
›Guides
Getting Started
Installation
Software Stack
Examples
Guides
Device
Tensor
Autograd
Optimizer
Model
ONNX
Distributed Training
Time Profiling
Half Precision
Development
Download SINGA
Build SINGA from Source
How to Contribute Code
How to Contribute to Documentation
How to Prepare a Release
Git Workflow
EditOptimizer
SINGA 支持各种流行的优化器，包括动量随机梯度下降、Adam、RMSProp 和 AdaGrad 等。
对于每一种优化器，它都支持使用衰减调度器来安排不同时间段的学习率。优化器和衰减调
度器包含在singa/opt.py中。
创建一个优化器
带动量的 SGD
# define hyperparameter learning rate
lr = 0.001
# define hyperparameter momentum
momentum = 0.9
# define hyperparameter weight decay
weight_decay = 0.0001
from singa import opt
sgd = opt.SGD(lr=lr, momentum=momentum, weight_decay=weight_decay)
RMSProp
# define hyperparameter learning rate
lr = 0.001
# define hyperparameter rho
rho = 0.9
# define hyperparameter epsilon
epsilon = 1e-8
# define hyperparameter weight decay
weight_decay = 0.0001
from singa import opt
sgd = opt.RMSProp(lr=lr, rho=rho, epsilon=epsilon, weight_decay=weight_decay)
AdaGrad
# define hyperparameter learning rate
lr = 0.001
# define hyperparameter epsilon
epsilon = 1e-8
# define hyperparameter weight decay
weight_decay = 0.0001
from singa import opt
sgd = opt.AdaGrad(lr=lr, epsilon=epsilon, weight_decay=weight_decay)
Adam
# define hyperparameter learning rate
lr = 0.001
# define hyperparameter beta 1
beta_1= 0.9
# define hyperparameter beta 2
beta_1= 0.999
# define hyperparameter epsilon
epsilon = 1e-8
# define hyperparameter weight decay
weight_decay = 0.0001
from singa import opt
sgd = opt.Adam(lr=lr, beta_1=beta_1, beta_2=beta_2, epsilon=epsilon, weight_decay=weight_decay)
创建一个衰减调度器
from singa import opt
# define initial learning rate
lr_init = 0.001
# define the rate of decay in the decay scheduler
decay_rate = 0.95
# define whether the learning rate schedule is a staircase shape
staircase=True
# define the decay step of the decay scheduler (in this example the lr is decreased at every 2 steps)
decay_steps = 2
# create the decay scheduler, the schedule of lr becomes lr_init * (decay_rate ^ (step // decay_steps) )
lr = opt.ExponentialDecay(0.1, 2, 0.5, True)
# Use the lr to create an optimizer
sgd = opt.SGD(lr=lr, momentum=0.9, weight_decay=0.0001)
使用模型 API 中的优化器
当我们创建模型时，我们需要将优化器附加到模型上：
# create a CNN using the Model API
model = CNN()
# initialize optimizer and attach it to the model
sgd = opt.SGD(lr=0.005, momentum=0.9, weight_decay=1e-5)
model.set_optimizer(sgd)
然后，当我们调用模型时，它会运行利用优化器的 train_one_batch 方法。
因此，一个迭代循环优化模型的例子是：
for b in range(num_train_batch):
# generate the next mini-batch
x, y = ...
# Copy the data into input tensors
tx.copy_from_numpy(x)
ty.copy_from_numpy(y)
# Training with one batch
out, loss = model(tx, ty)
Last updated on 4/7/2020
← AutogradModel →
创建一个优化器
创建一个衰减调度器
使用模型 API 中的优化器
DocsGetting StartedGuidesAPI ReferenceExamplesDevelopment
CommunityUser ShowcaseSINGA HistorySINGA TeamSINGA NewsGitHub
apache/singa
Follow @ApacheSINGA
Apache Software FoundationFoundationLicenseSponsorshipThanksEventsSecurity
Copyright © 2026
The Apache Software Foundation. All rights reserved.
Apache SINGA, Apache, the Apache feather logo, and
the Apache SINGA project logos are trademarks of The
Apache Software Foundation. All other marks mentioned
may be trademarks or registered trademarks of their
respective owners.