This is the repo for the OpenD/I Hackathon 2023, HyperCycle challenge 1/2. While the primary development of this repo was designed for the decentralized incremental learning challenge, the anonymity protections in the solution (via Tor support) may qualify for the privacy focus of challenge 1.
During the first half of the hackathon, my focus was on making the network mode functional, and during the second half, it was to create example genetic algorithms applied to neural architectures to show that they can rediscover classic findings in deep learning. However, while having issues with the skip connection task, I had the idea for the GAS algorithm, and my focus shifted towards that instead.
As such, most of the genetic algorithm tasks were implemented in the single task mode instead of the network mode (with clients and servers), so they could be quicker to develop and test. The reference implementation for the decentralized challenge is in the task01 task and is referenced below in the network mode section.
Before running, be sure to have git, python3, pip, and virtualenv installed. Also make sure you have mongodb installed and are running a database on port 27017 (default).
sudo apt install python3 python3-dev virtualenv
Clone the repo:
git clone https://github.com/amir-f34/SublimeAlgo
To run a server:
First, start a mongodb instance, and then run:
virtualenv venv
source ./venv/bin/activate
cd SublimeAlgo
pip install -r requirements.txt
cd server
uvicorn serve:app --reload
To create a new task (for example: task09):
1.) Add task09 to tasks/
2.) Add a corresponding task manager to task_manager.py for genetic algorithm specifics.
3.) Run python3 create_task.py task09
4.) Run/Restart the task manager: python3 task_manager.py. Be sure to add the task to the manager loop in main().
To run the example (task01) in the demo, just run python3 task_manager.py
To create a trusted key (for reputation bootstrapping)
python3 create_trusted.py <Public Key>
To run a client:
virtualenv venv
source ./venv/bin/activate
cd SublimeAlgo
pip install -r requirements.txt
cd client
python3 sublime_cli.py ...
For help options:
python3 sublime_cli.py -h
To specify a different config file, or create a new one if it doesn't exist:
python3 sublime_cli.py -f subl2.conf
To show current tasks:
python3 sublime_cli.py -s
To download a task (with autoverify off):
python3 sublime_cli.py -t task01-18ed0f86
To mark a task as verified:
python3 sublime_cli.py -v task01-18ed0f86
To run a verified task:
python3 sublime_cli.py -t task01-18ed0f86
You can also use the CLI with the API server:
uvicorn client_serve:app --reload --port=8001
And then access it with:
curl -XPOST http://localhost:8001/sublime_cli -d '{"args": ["-s"]}'
Install the requirements as in the network mode, and then run
python3 task_single.py
followed by either --task1 --task4 --task5 or --task8
Task1 is a basic string count example, while task4 is the activation function experiment. Task5 is the attempt made at adding in skip-connections using different chromosomes for the different branching of the network (which was unfinished), while Task8 is the genetic algorithm and genetic A* implementations.
Amir F.
0x1D7e4845f19c801651280CBDD84d70ea511bF954