andreapisa9 / bayesian-elm-search

As described in "Towards Full On-Tiny-Device Learning: Guided Search for a Randomly Initialized Neural Network"

Bayesian ELM Search

---

License

This work is licensed under Creative Commons BY-NC-ND 4.0. Please read the license text carefully before handling the code.

Attribution

If you choose to utilize our work for your research, please cite the following in your work:

@article{pau2024towards,
  title={Towards Full Forward On-Tiny-Device Learning: A Guided Search for a Randomly Initialized Neural Network},
  author={Pau, Danilo and Pisani, Andrea and Candelieri, Antonio},
  journal={Algorithms},
  volume={17},
  number={1},
  pages={22},
  year={2024},
  publisher={MDPI}
}

---

How to Use

Install the Required Libraries

The experiments were run on a conda environment. To reproduce them, please install anaconda or miniconda and create an empty environment using the following command:

conda create -n bayesian-elm-search python=3.9.12

After that, please activate the environment using the following command:

conda activate bayesian-elm-search

Finally, move to the project folder named src and install the required libraries using pip with the following command:

pip install -r requirements.txt

Download the CIFAR-10 Dataset

Please download in a folder named 'CIFAR-10' the dataset from its original source. Be careful to download the python version (161 MB). In the same folder, download the Python script perf_sample_idxs.py from the MLCommons Tiny repo.

Run the Experiments

Experiments are run using the following Python files:

• FE_bayes_GP.py, which uses Type 1 neural topology and Gaussian Processes as surrogate model;
• FE_bayes_RF.py, which uses Type 1 neural topology and Random Forests as surrogate model;
• FE_bayes_RF_newtopology.py, which uses Type 2 neural topology and Random Forests as surrogate model.

Within the Python scripts, please set the WORKING_DS string to be equal to the name of the dataset you may want to experiment with (i.e. 'MNIST' or 'CIFAR-10'). After that, launch the scripts one at a time using the following command:

python3 SCRIPTNAME.py > OUTPUT_SCRIPTNAME.txt

The results will be incrementally written on text files within the same folder. Please note that FE_bayes_GP.py automatically saves the results within a JSON file called logs.log.json, while the other two print the results on standard output.