EvoDENSS

EvoDENSS stands for Evolution of DEep Networks through Self Supervision. EvoDENSS performs Neuro-Evolution by evolving the structure of the networks and optimiser aspects (the optimiser instance to use and relevant hyperparameters). Being inspired on Fast-DENSER, it uses self-supervised learning to train the generated networks in order to measure their fitness.

Installing

In order to run EvoDENSS, one needs to install the relevant dependencies. There are two ways to install the framework:

1. Conda

A conda environment can be created from an exported yml file that contains all the required dependences:

conda env create -f environment.yml

After the environment is created, just activate it in order to be able to run your code:

conda activate evodenss

2. pip

Alternatively, you can use the requirements.txt file, but you will be on your own to install cudatoolkit and other libraries that might be required to enable GPU acceleration.

pip install -r requirements.txt

Note: Installing EvoDENSS as a Python library is not yet supported

Running EvoDENSS

• In order to run fast EvoDENSS, you need to run the evodenss module as a script:

python3 -m evodenss.main \
    -d <dataset_name> \
    -c <config_path> \
    -g <grammar_path> \
    -r <#run>

Example:

python3 -m evodenss.main \
    -d mnist \
    -c example/example_config.yaml \
    -g example/example.grammar \
    --run 0 \
    --gpu-enabled

In case several seeds are needed to be run, that can be done with Bash:

for i in {7..9}; do \
python3 -m evodenss.main \
-d cifar10 \
-c config_files/bt_10.yaml \
-g grammars/bt.grammar \
-r $i --gpu-enabled; \
done

Externally to the code itself, two main files are required to execute any run.

1. A grammar that shapes the search space by setting the possibilities within each macro block.
2. A configuration file that sets miscellaneous parameters that affect the outcome of the evolutionary run. These can se related with the evolutionary process itself, or the networks that are generated.

Testing

Unit tests can be executed via pytest:

pytest tests

In case one wants to do it with coverage report:

coverage run --source evodenss -m pytest -v tests
coverage report

Command-line flags

• -c/--config-path: Sets the path to the config file to be used;
• -d/--dataset-name: Name of the dataset to be used. At the moment, mnist, fashion-mnist, cifar10 and cifar100 are supported.
• -g/--grammar-path: Sets the path to the grammar to be used;
• -r/--run: Identifies the run id and seed to be used;
• --gpu-enabled: When used, it enables GPU processing.

Framework modes

EvoDENSS can be run using the supervised learning mode (to mimic Fast-DENSER behaviour) and the self-supervised learning mode. In the case of self-supervised learning there are a few variations that influence which components are targeted by Evolutionary Computation.

1. Supervised learning mode

2. Self-Supervised learning mode

2. Self-Supervised learning mode with evolvable projector

TBD in an upcoming version

Versioning

Versioning is done according to the Semantic Versioning guidelines. Given a version number MAJOR.MINOR.PATCH, increment the:

MAJOR version is incremented whenever a breaking change is added. MINOR version is incremented whenever new features are added with backwards compatibility. PATCH version is incremented whenever bug fixes are done with backwards compatibility.

Numbers are updated by semi-automatically by autosemver library, according to the commit messages that are passed. For more information about this, please read autosemver documentation.