EVO X

Working with Module in EvoX

Working with Module in EvoX

A module is a fundamental concept in programming that refers to a self-contained unit of code designed to perform a specific task or a set of related tasks.

This notebook will introduce the basic module in EvoX: ModuleBase.

Introduction to Module

In the tutorial, we have mentioned the basic running process in EvoX:

Initiate an algorithm and a problem -- Set an monitor -- Initiate a workflow -- Run the workflow

This process requires four basic class in EvoX:

  • Algorithm
  • Problem
  • Monitor
  • Workflow

It is necessary to provide a unified module for them. In EvoX, the four classes are all inherited from the base module — ModuleBase.

Module base

ModuleBase class

The ModuleBase class is inherited from torch.nn.Module.

There are many methods in this class, and some important methods are here:

MethodSignatureUsage
__init__(self, ...)Initialize the module.
load_state_dict(self, state_dict: Mapping[str, torch.Tensor], copy: bool = False, ...)Copy parameters and buffers from state_dict into this module and its descendants. It overwrites torch.nn.Module.load_state_dict.
add_mutable(self, name: str, value: Union[torch.Tensor | nn.Module, Sequence[torch.Tensor | nn.Module], Dict[str, torch.Tensor | nn.Module]]) -> NoneDefine a mutable value in this module that can be accessed via self.[name] and modified in-place.

Role of Module

In EvoX, the ModuleBase could help to:

  • Contain mutable values

​ This module is an object-oriented one that can contain mutable values.

  • Support functional programming

​ Functional programming model is supported via self.state_dict() and self.load_state_dict(...).

  • Standardize the initialization:

​ Basically, predefined submodule(s) which will be ADDED to this module and accessed later in member method(s) should be treated as “non-static members”, while any other member(s) should be treated as “static members”.

​ The module initialization for non-static members are recommended to be written in the overwritten method of setup (or any other member method) rather than __init__.

Usage of Module

Specifically, there are some rules for using ModuleBase in EvoX:

Static methods

Static methods to be JIT shall be defined like:

# One example of the static method defined in a Module

@jit
def func(x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
    return x + y

Non-static Methods

If a method with python dynamic control flows like if were to be used with vmap, please use torch.cond to explicitly define the control flow.

← Back to Documentation