Code for A Trainable Lossless Syntax Tree Generator

Presented by:

• Chia-Yi Su
• Robert Wallace
• Juliana Gonzalez
• Vijayanta Jain
• Sepideh Ghanavati
• Collin McMillan

This repository contains all the code and detailed instructions for a tool to generate lossless syntax trees from source code using a language mode and repair syntax errors in our HuggingFace Automatic Program Comprehension Lab hub.

Quick link

• To-do list
• Error Correction in Zero-Shot Setting
• Finetuning
• Metrics
• Dataset
• Retrain from scratch

To-do list

To set up your local environment, run the following command. We recommend the use of a virtual environment for running the experiments.

pip install -r requirements.txt

• If you want to generate the syntax trees from source code or correct syntactic eorrs in zero-shot setting with our models, please see Error Correction in Zero-Shot Setting.
• If you want to finetune a model to fix the syntactic bug using our processed and tokenized dataset, please see Finetuning
• If you want to retrain the model from scratch, please see Retrain from scratch

Error Correction in Zero-Shot Setting

Step 1: Download dataset and models

Please downlaod the fundats.tar.gz file in our Hugginface dataset repo and download the ckpt_base.pt file in the model repo and place all of the files in fundats.tar.gz in /nublar/datasets/jm52m/ and palce the model file in jmsrcml

Step 2: Syntax tree generation

CUDA_DEVICE_ORDER='PCI_BUS_ID' CUDA_VISIBLE_DEVICES='1' OMP_NUM_THREADS=2 time torchrun --rdzv-backend=c10d --rdzv-endpoint=localhost:4111 --nnodes=1 --nproc_per_node=1  sample_srcml.py --out_dir=jmsrcml --temperature=0.001 --prediction_outdir=srcml_prediction_new --checkpoint_filename=ckpt_base.pt

--out_dir: directory of the model for inference
--prediction_outdir: name of the directory of the prediction file
--checkpoint_filename: the filename of the inference model
--q90codefile: filename of the function
--q90codetestfidfile: filename of the funtion id

Step 3: Code generation from syntax tree

python3 decoded_srcml.py 

--srcml_dir: directory of syntax tree files
--q90codefile: function files
--q90testfidfile: filename of the function id
--decoded_code_file: filename of decoded code

Finetuning

These steps will show you how to fine-tune the model to fix the syntax errors and make the inference by using the model that you finetune

Step 1: Download the finetuning dataset

Please download bin.tar.gz in our Hugginface repo and put train.bin and val.bin to the same dir as --dataset in config/finetune_autorepair.py, which is data/autorepair for now.

Step 2: Download the models for finetuning

Please download the checkpoint files named ckpt.pt in our Hugginface repo for finetuning and put the checkpoint to the same dir as --out_dir in config/finetune_autorepair.py.

Step 3: Finetuning model

CUDA_DEVICE_ORDER='PCI_BUS_ID' CUDA_VISIBLE_DEVICES='0' OMP_NUM_THREADS=2 time torchrun --rdzv-backend=c10d --rdzv-endpoint=localhost:4000 --nnodes=1 --nproc_per_node=1  train.py config/finetune_autorepair.py --outfilename=ckpt.pt

Step 4: Inference

CUDA_DEVICE_ORDER='PCI_BUS_ID' CUDA_VISIBLE_DEVICES='0' OMP_NUM_THREADS=2 time torchrun --rdzv-backend=c10d --rdzv-endpoint=localhost:4000 --nnodes=1 --nproc_per_node=1 sample_autorepair.py config/finetune_autorepair.py --prediction_filename=predict_autorepair_srcml.pkl --outfilename=ckpt.pt

Metrics

We provide scripts for calculating the metrics to evaluate the srcml and bug fixing rate bellow.

Merics for evaluating syntax tree

python3 eval_srcml.py

--srcml_dir: syntax tree prediction directory
--q90codefile: function file
--q90testsrcmlfile: reference syntax tree file
--q90testfidsfile: function id file
--q90decodedcodefile: decoded code file

Merics for evaluatig finetuning bug fixing rate

python3 autorepair_base_fix_rate.py

--reference_code_file: filename of the reference code
--prediction_file: filename of the prediction code

Merics for evaluatig zero-shot bug fixing rate

python3 srcml_bug_fix_rate.py

srcml_dir: directory of syntax tree files
q90testfidsfile: filename of funtion id
bug_code_file: filename of the funtions with syntax errors
q90codefile: filename of reference code

Dataset

We also release all of our raw datasets for finetuning in our Hugginface repo and the scripts for compiling the raw data to bin files in this Github repo. Before running the command, please create three dir: pkls, bins, and tmp. Then, you can simply run the following command to generate train.bin and val.bin.

python3 data/autorepair/prepare_fc_raw.py

--q90trainfids-file: filename of training function id
--q90testfids-file: filename of test function id
--q90valfids-file: filename of val function id
--fundats-file: filename of function
--train-fundats-file: filename for the function with the syntax error for training
--val-fundats-file: filename for the function with the syntax error for val

Retrain from scratch

Step 1: Download the dataset

Please download train.bin.gz and val.bin.gz in our Hugginface repo and extract and put those files to the same dir as --dataset in config/pretraining.py, which is data/pretrain for now.

Step 2: Retrain model

CUDA_DEVICE_ORDER='PCI_BUS_ID' CUDA_VISIBLE_DEVICES='0' OMP_NUM_THREADS=2 time torchrun --rdzv-backend=c10d --rdzv-endpoint=localhost:4000 --nnodes=1 --nproc_per_node=1  train.py config/pretraining.py