pulsatrix 1.1.0
Loading...
Searching...
No Matches
pulsatrix::ParamGroup Struct Reference

Optimizer settings for the parameters selects picks. More...

#include <param_groups.hpp>

Public Attributes

std::string name
 For error messages and logs.
 
ParamSelector selects
 
float learning_rate
 
float weight_decay = 0.0f
 L2 penalty: the step uses grad + weight_decay * value, as PyTorch's SGD and Adam do.
 

Detailed Description

Optimizer settings for the parameters selects picks.

Note
An optimizer puts each parameter in the first group that selects it, and every parameter no group selects in its default group (the optimizer's own learning rate and weight decay). learning_rate and weight_decay may be changed between steps, which is how a learning-rate schedule adjusts a group.

Member Data Documentation

◆ learning_rate

float pulsatrix::ParamGroup::learning_rate

◆ name

std::string pulsatrix::ParamGroup::name

For error messages and logs.

◆ selects

ParamSelector pulsatrix::ParamGroup::selects

◆ weight_decay

float pulsatrix::ParamGroup::weight_decay = 0.0f

L2 penalty: the step uses grad + weight_decay * value, as PyTorch's SGD and Adam do.


The documentation for this struct was generated from the following file: