Autoencoder Program Synthesis: A Tree-based VAE-RNN as a tool for generating new C++ programs.
- Make sure to have PyTorch installed: https://pytorch.org/get-started/locally/.
- Make sure to install clang LLVM https://github.com/llvm/llvm-project/releases/tag/llvmorg-12.0.1, for windows make sure to add LLVM to the system path
- Install clang-format for automatic formatting of the generated code, run:
apt-get install clang-format - To install this as a package run:
pip install git+https://github.com/sander102907/autoencoder_program_synthesis
- Download a pretrained model checkpoint from here
- Unzip the contents and place the checkpoint in a folder
- Locate the libclang.so library file on your computer and save the path
Then you can simply import the package in your python code and encode C++ programs to latent vectors and decode latent vectors back to strings. The encoding function takes a program represented as a string and the decoder takes a torch tensor, and some sampling parameters: temperature, top_k and top_p (see this paper for more information). Use temperature=0, top_k=0 and top_p=0 for greedy decoding. The pretrained model that is provided in the set up works with latent vectors of size 150, hence the decoder takes a torch tensor of shape [X, 150] where X can be any number of programs you want to generate. An example of a reconstruction:
from autoencoder_program_synthesis.t2t_vae import Tree2Tree
libclang_path = '/usr/lib/x86_64-linux-gnu/libclang-6.0.so.1'
checkpoint_path = 'checkpoints/1/'
tree2tree = Tree2Tree(libclang_path, checkpoint_path)
program = """
#include <bits/stdc++.h>
using namespace std;
int main() {
ios::sync_with_stdio(false);
cin.tie(NULL);
int n;
cin >> n;
map<int,int> cn;
int ans = 0;
for(int i = 0; i < n; i++) {
int x;
cin >> x;
cn[x]++;
ans = max(ans,cn[x]);
}
cout << ans;
return 0;
}"""
z, decl_names = tree2tree.encode(program)
reconstructed_program = tree2tree.decode(z, temperature=0.7, top_k=40, top_p=0.9, declared_names=decl_names)
Model training/testing/finetuning