Hierarchically Organized Latent Modules for Exploratory Search in Morphogenetic Systems

Project Website

Mayalen Etcheverry, Clément Moulin-Frier, Pierre-Yves Oudeyer
Flowers Team
Inria, Univ. Bordeaux, Ensta ParisTech (France)

This repository hosts the source code to reproduce the results presented in the paper Hierarchically Organized Latent Modules for Exploratory Search in Morphogenetic Systems.

Step 0: Installation of the conda environment

1. If you do not already have it, please install Conda
2. Create holmes conda environment: conda create --name holmes python=3.6
3. Activate holmes conda environment: conda activate holmes
4. Install the required conda packages in the environment (one by one to deal with dependencies errors): while read requirement; do conda install --yes $requirement --channel default --channel anaconda --channel conda-forge --channel pytorch; done < requirements.txt
5. Install the required pip packages in the environment (including the provided packages exputils, autodisc and goalrepresent): pip install -e .

Step 1: Reproduce paper results

To reproduce a figure from the paper, please do the following:

	cd reproduce_paper_figures
	jupyter notebook

Then open the notebook corresponding to the figure you want to reproduce and run all the cells.

Step 2: Reproduce paper experiments

You can find the python code to reproduce all the experiments from the paper (main and supplementary), by going in the experiments folder:

experiments/
├── IMGEP-BC-BetaVAE
├── IMGEP-BC-EllipticalFourier
├── IMGEP-BC-LeniaStatistics
├── IMGEP-BC-PatchBetaVAE
├── IMGEP-BC-SpectrumFourier
├── IMGEP-BetaTCVAE
├── IMGEP-BetaVAE
├── IMGEP-BigVAE
├── IMGEP-HOLMES
├── IMGEP-HOLMES_no_connection
├── IMGEP-HOLMES_only_gfi_c
├── IMGEP-HOLMES_only_lf_c
├── IMGEP-HOLMES_only_lfi_c
├── IMGEP-HOLMES_only_recon_c
├── IMGEP-HOLMES (SLP)
├── IMGEP-HOLMES (TLP)
├── IMGEP-SimCLR
├── IMGEP-TripletCLR
├── IMGEP-VAE
└── Random Exploration

Each folder IMGEP-X corresponds to one algorithm variant presented in the paper, and is organized as follow:

experiments/IMGEP-X/
├── calc_statistics_over_repetitions.py ├── calc_statistic_space_representation.py
├── repetition_000000
├── repetition_000001
├── repetition_000002
├── repetition_000003
├── repetition_000004
├── repetition_000005
├── repetition_000006
├── repetition_000007
├── repetition_000008
├── repetition_000009
└── statistics

Each subfolder repetition_00000i corresponds to one repetition directory (seed=i) for each algorithm variant, and is organized as follow:

experiments/IMGEP-X/repetition_00000i/
├── calc_holmes_RSA_per_repetition.py
├── calc_statistics_per_repetition.py
├── calc_temporal_holmes_RSA_per_repetition.py
├── calc_temporal_RSA_per_repetition.py
├── experiment_config.py
├── neat_config.cfg
├── run_experiment.py
└── statistics

We already provide all the necessary saved results from our experiments in the statistics subfolders. The figures of the main paper (Cf Step 1) are generated by loading the saved results.

If you wand to regenerate those results, two steps are needed:

1. Run all the individual training experiments. For instance, to run repetition 0 of IMGEP-HOLMES you should do the following:

	cd experiments/IMGEP-HOLMES/repetition_000000/
	conda activate holmes
	python run_experiment.py

2. Run statistics per experiment and/or per repetition, once the training is completed. For instance, for regenerating all the statistics used in IMGEP-HOLMES you should do the following:

	cd experiments/IMGEP-HOLMES/
	conda activate holmes
	python calc_statistic_space_representation.py
	python calc_statistics_over_repetitions.py 
	cd repetition_000000/
	python calc_holmes_RSA_per_repetition.py

Notice: Please note that each individual training experiments needs a long time to train (between 20 and 35 hours with one GPU), we therefore recommand to run them if possible on a cluster and in parallel. For this purpose, each python script <script_name>.py is accompanied by a run_<script_name>.slurm to run the code on a cluster using the SLURM job manager.

Acknowledgement

The exputils and autodisc packages used in our code builds upon flowersteam's packages developped by Chris Reinke.