PyTorch-CNS

An implementation of Generalized Compressed Network Search for PyTorch models.

This creates a genome per layer, rather than a single one for the entire model as is described in the paper.

There are two optimizers:

• Asynchronous Gene Pool: a master list of genomes, sorted by fitness, is worked on by a pool of agents which draw fit genomes for mutation and evaluation.
• Synchronous Score Swap: each worker maintains a full copy of all genomes and only fitness scores are exchanged.

Installation

Install python package with pip install pytorch-cns

Install redis which is used as the datastore. The examples expect a redis instance listening on localhost at the default port. You can change this by passing a JSON dict of StrictRedis kwargs via the --redis-params command line option.

AI Gym Examples

• aigym.py: CartPole!
• atari.py: Atari ram-based games
• atari_pix.py: Atari pixel-based games
• atari_pixrnn_gpa.py: Atari pixel-based games with a recurrent neural network, using the asynchronous gene pool optimizer.
• atari_pixrnn_ss.py: Atari pixel-based games with a recurrent neural network, using the synchronous

Install additional requirements: pip install gym atari_py box2d

To run a pool of workers with default settings simply run the python file (e.g. python atari_pix.py). If you make any changes to the hyperparameters you'll to use the --clear-store flag which deletes the old gene pool upon start. Use --num-agents to customize the number of child processes spawned.

Invoke the example with python atari_pix.py --render --best to run the simulation with the fittest genome. This can be done at the same time as the workers are running to monitor progress.

Image generation

These do not converge on anything at the moment. Maybe, if you are a real GANimal, you can find the right configuration. The code here is hacked together from the pytorch example repo.

cnsdcgan.py: DCGAN adapted from the PyTorch DCGAN example. Attempts to train both the discriminator and generator with compressed network search.

vggmse.py: An autoencoder which uses VGG16 to calculate the loss.