Deep Reinforcement Learning for Efficient Neural Architecture Search (ENAS) in PyTorch, i.e., AutoML. Code based on the paper https://arxiv.org/abs/1802.03268
How to run the pipeline:
-
Clone the whole repository
git clone https://github.com/RualPerez/AutoML.git
-
Create virtualenv
virtualenv -p /usr/bin/python3.7 virtualenv/AutoML
-
Activate virtualenv
source virtualenv/AutoML/bin/activate
-
Install libraries
pip3 install -r requirements.txt
-
Run the main script, for instance:
python3 main.py --num_episodes 5 --batch 5 --possible_hidden_units 1 4
Note that you can get a help of how to run the main script by:
python3 main.py -h
Once the whole steps have run successfully, the next times you only need to run the last step 5.
Output: The main script saves the trained policy/controller net as policy.pt
File / Folder | Description |
---|---|
main.py | Main script with runs the AutoML experiment |
*.py | An auxiliar script necessary to run main.py, its detailed description has been written as python documentation |
Policy_Gradient_AutoML.ipynb | Jupyter Notebook designed to help users to understand how this project has been developed |
article.pdf | Basic article that describes the principles of this project (theory-related). Here it can be found the results. |
requirements.txt | Version of the python libraries necessary to run the main script |
images/ | Images used for the article |