The code for the experiments generating SCA CNNs can be found in the metaqnn folder.
The definitions for the experiments are in the models folder.
To run an experiment, use python -m metaqnn.main <model>, where model can be any of the folders in the models folder.
To generate an overview of the results and to create the scatter plots, use python -m metaqnn.display_results -h for instructions.
To generate the GE graph or use the simple ensemble, use python -m metaqnn.plot_top_ges -h for instructions.
requirements.txt includes all requirements including the exact dependencies used, while requirements.minimal.txt only includes the explicitly installed requirements (generated with pip-chill)
NB: at most tensorflow 2.1, cuda 10.1 and cudnn 10.1-7.6.0.64 are required. Higher tensorflow versions have some breaking changes that cause the code crash (mostly the One-Cycle LR code).
This code is partially based on:
Designing Neural Network Architectures Using Reinforcement Learning
Bowen Baker, Otkrist Gupta, Nikhil Naik, Ramesh Raskar
International Conference on Learning Representations, 2017
The source code of which can be found on Github under the MIT License.