moduleimplant 0.2.0

A custom module implant system for Ultralytics models

Module-Implant

This project can help you implant 3rd party / custom pytorch nn.module components into Ultralytics library. After that, you can simply modify the yaml file, which defines your network structure, with new components.

1. Install

There are 2 ways to install this project. You can simply download / clone the project and follow the manual instructions below, or try:

pip install moduleimplant

2. Usage

Auto setup

Module-Implant need some manual operations to modify Ultralytics since Ultralytics does not provide APIs for custom modules and using if-else to select inner modules.

For auto setup, try:

moduleimlpant modify

for modifying Ulralytics pack;

moduleimlpant de-modify

for cancle modification before uninstallation.

manual setup

Paste moduleimplant.py file where task.py located, modify task.py as following:

1. import

   from .moduleimplant import ModuleImplant
   

2. update globals() during parse_model()

   globals().update(ModuleImplant.get_third_party_modules_dict())
   # insert this before 'for i, (f, n, m, args) in enumerate...'
   

3. add this elif within others in the for statement that mentioned above

   elif m in ModuleImplant.get_third_party_modules():
       c1, c2, args = ModuleImplant.parse_third_party_module(ch, f, n, m, args)
   

Then, modify the paralled __init__.py file as following:

1. add this piece of code

   from .moduleimplant import (
       ModuleImplant
   )
   

2. add "ModuleImplant" to __all__

prepare custom modules

You also need a parser func to parse arguments provided by .yaml file. Implement a func called 'yaml_args_parser' in your custom module like this:

@staticmethod
def yaml_args_parser(ch, f, n, m, args):
    return ch[f], ch[f], [ch[f], *args]

tasks.py will pass 5 arguments for your custom module:

• ch: a list that recorded all layers' output channel count
• f: previous layer's serial number which indicates where this layer's data comes from
• n: assume that your layer should be a serial of modules, n represents the repetition times
• m: definition form of this module
• args: custom arguments after aboves

tasks.py needs 3 return values:

• c1: layer's input channel count
• c2: layer's output channel count
• args: all arguments that module's initial function requires