prog-synth-SAC

Lucas Kabela

A research project investigating applications of Soft Actor Critic in Neural Program Synthesis

Final Writeup

References:

This project is largely following the work of Robustfill by Delvin et al. 17

DSL and base model was largerly derived from here

SAC imlementation derived from here

Getting Started

Prerequisites

The following base packages were used to run this repository:

• Python - 3.6+
• Numpy - 1.16.4+
• PyTorch - Lastest (1.2.0+)

First Steps

This repository contains code for training a model for neural program synthesis. We provide supervised learning and reinforcement learing algorithms REINFORCE and SAC. To train a model, run python train.py [--sac] [--reinforce] and to evaluate python train.py --eval -c --checkpoint_filename [policy to evaluate]

See train.py for more command line arguments related to hyperparameters. Alternatively, we have provided a notebook, train.ipynb which requires the code in the src folder to run. This code, with supervised training from scratch should reach a loss of 4 within minutes, and below 2 in about 30 hours with hyperparameters provided

Repository Structure

.
├── model_zoo              # Pretrained models used to produce results
|
├── results                # Results of experiments in csv
|
├── src                    # source code for the project
|
├── writeups               # Project proposal and final report
|
├── LICENSE
| 
└── README.md

Src

Contains the code required for training the models and running experiments, as well as executable notebooks.

Training in Colab:

Note, this program has long training times, so if you are running in colab, to avoid disconnection from inactivity and take advantage of full training time, please add the following to the console:

function ClickConnect(){
    console.log("Clicking");
    document.querySelector("colab-connect-button").click()
}
setInterval(ClickConnect,60000)

Results

Contains the raw data from experiments

Writeup

This folder contains the project proposal and final report/writeup describing the methodology, intuitions, related works and results.

License:

This project is licensed under the terms of the MIT license.