This is a re-implementation of the first neural architecture search paper: https://arxiv.org/abs/1611.01578.
We test our implementation on three setting:
- customized toy environment, where the dataset is a multi-modal Gaussian;
- A 3-layer fully-connected network on Google speech commands dataset;
- A simple CNN on Google speech commands dataset.
To run the code, simply go to the toy-model folder and run
python pg.py
Instructions on how to run the google speech commands dataset will be updated soon. However, you have all the scripts needed in this repo. Please take a look on our report: cs234-final-report.pdf in this repo for more details.