Benchmarking learning efficiency in deep reservoir computing

This is the code to reproduce results from the paper

Benchmarking Learning Efficiency in Deep Reservoir Computing. Cisneros, H., Mikolov, T., & Sivic, J. (2022). 1st Conference on Lifelong Learning Agents, Montreal, Canada.

Re-run experiments

WARNING: Re-running all experiments might take a significant amount of time. Experiments in the paper were done on a cluster using GPUs and a lot of parallelism. The docker solution is particularly sub-optimal and will take a long time to run experiments.

An alternative to running all the experiments is to download the data directly:

wget https://data.ciirc.cvut.cz/public/projects/2022BenchmarkingLearningEfficiency/experiment_2022-07-13T15:32:50.tar

tar -xvf "experiment_2022-07-13T15:32:50.tar"

Running with poetry

The easiest way to run the experiments is to use poetry. First, clone the repo

git clone https://github.com/hugcis/benchmark_learning_efficiency.git

Then, run poetry install to create a virtual environment and install all the dependencies.

Then run:

./run_experiments.sh

Running in Docker

If you don't have or don't want to install poetry, you can build and install everything within a docker container. Just run the following from inside the repo:

docker build -t pypoetry_bledrc .
docker run -it --entrypoint=/bin/bash pypoetry_bledrc -i

This will open a bash tty within the docker container where you can run

./run_experiments.sh

Generate figures and tables

Once the data is generated or downloaded (make sure that you have the experiment_gpu and experiment_sgd folders), you can run jupyter notebooks in order to re-generate the figures and tables from the paper.

Just run

poetry run jupyter notebook

and open the two jupyter notebooks in the folder notebooks.