automatai's Introduction

AutomatAI

just for fun, and also because I don't like spending all the day tweaking hyperparameters, I'll use a genetic algorithm and later a particle swarm optimisation to create neural networks

How does it work ?

for now this is kinda spaghetti code. Depending the results I'll create later an API or a CL app

pro:

lazy creation of a NN
time consuming (or you need a lot of powerful hardware or small dataset and small data)

deps & run

python35 -m venv env

source env/bin/activate

pip install -r requirements.txt

# OR

pip install -r req_dev.txt

python main.py

Parameters

Folder (In, Checkpoints, Out)
Population size
RAM size
Number of GPU + DRAM
Data shape (In, Out/validation)
Methodology (supervised, semi-supervised)
Stop & monitoring parameters
- Loss or accuracy evolution (stop)
- Delta bewteen train and validation scores (stop)
- Maximum/minimum accuracy/loss (monitoring)
- Data usage (semi-supervised) (monitoring)
- Benchmark result ? (monitoring) / possibly highly time consuming
Benchmarks ? (how fast is it on a cpu / gpu / arm chip)

Hyper parameters selection

Each of those parameters can be choosen before launching the algorithm, or let the algorithm automaticaly use them in a generation

Loss function (optionnal)
Optimizer (optionnal)
Layers (optionnal, not planned yet)
Layer parameters:
- Kernel size
- Regularisers
- Ititializers
- number of connected points (FCL)
Data generators (optionnal)

Recommend Projects