TFix: Learning to Fix Coding Errors with a Text-to-Text Transformer

TFix is a state-of-the-art system for automatically fixing coding errors in programs. The key idea behind TFix is to leverage a large text-to-text Transformer pre-trained on natural languages. This design allows TFix to apply a knowledge transfer between natural and programming languages. In addition to that, TFix is fine-tuned jointly on 52 different error types, which allows it to learn typical patterns across various error types together.

Paper

The paper can be found under this and this link.

If you find our paper useful, please cite:

@inproceedings{Berabi2021TFixLT,
  title={TFix: Learning to Fix Coding Errors with a Text-to-Text Transformer},
  author={Berkay Berabi and Jingxuan He and Veselin Raychev and Martin T. Vechev},
  booktitle={ICML},
  year={2021}
}

Citation for other formats like Endnote, APA and a more detailed BibTeX citation can be found here at the very bottom.

Setup

To use TFix, you need Python 3. First create a virtual environment and install the dependencies:

python3 -m venv venv_tfix
source venv_tfix/bin/activate
pip install -r requirements.txt

Note that you may need to install torch, torchvision and torchtext according to your GPU and CUDA version.

Linters Setup

We provide our configuration and javascript files to run the ESLint with TFix's configuration and versions. Please note that it is crucial to have the same setup as otherwise the reported bugs can differ with varying configurations and versions.

ESLint with default config

cd linters/javascript/eslint_eslint
npm ci
npx webpack

ESLint with repo specific config

cd linters/javascript/eslint_repo_specific/
npm ci
npx webpack

From now, you should be able to run linters with both configuration. There are two options to run them. You can either directly provide a path to the file or you can provide a line containign a json line with certain fields. The latter can be more useful when you want to run it programatically on many files by using pipes. Please see the example commands below for more information.

// running eslint with default config on a file. The commands are the same for the repo_specific config. Just switch the directory accordingly.
cd linters/javascript/eslint_eslint

Option 1: Provide a path to the file you want to lint.

node src/index.js my_file_for_linting.js

Option 2: Provide a json as input. When you use this option, the running code will wait for an input from the standard input. So, the code hangs until you provide the input. The code expectsa single line that is a json line.

node src/index.js -- - JSON
// now provide a json like wit h the following format.
[{"code":"var x = 5;\nvar y = 11;\n","file_name":"some/folder/some_file.js"}]

Dataset and Models

The dataset and trained models used in our experiments are available under this link. Download and unzip them into the same directory as the code. We used the model named t5large for TFix.

The dataset contains a metadata for each data point (buggy-fixed code pair). The metadata is described in the table below.

Fields	Description
source_code	The code patch around the error that is fed to the model.
target_code	The corresponding fixed version of the buggy patch. It is fed to the model as target output.
repo	The repository from which the sample was extracted. Ignore the local path /data/all/data
source_changeid	The commit id of the file with the bug
target_changeid	The commit id of the file with the fix
source_filename	The name of the file with the bug
target_filename	The name of the file with the fix
source_file	It contains a larger patch around the bug that was used for debugging purposes. The name is misleasind since it does not contain the whole file.
target_file	It contains a larger patch around the fix that was used for debugging purposes. The name is misleasind since it does not contain the whole file.
linter_report	Contains the information reported by detector. It hals its own sub-fields like rule_id (error type) message (error message), line_begin and more
warning_line	The line of code on which the errror was reported by the detector.
instructions	It contains a list of text edit operation that explains the diff between source_code and target_code

Configuration

The main scripts are tfix_training.py and tfix_testing.py.

The tfix_training.py has the following arguments:

Required arguments:

 -mn --model-name   Name of the model. Choices: [t5-small, t5-base, t5-large, t5-3b, t5-11b]

Optional arguments:

Parameter	Default	Description
-e --epochs	1	Number of epochs to fine-tune the model
-bs --batch-size	1	Batch size to be used in fine-tuning
-lr -–learning-rate	1e-4	The initial learning rate for fine-tuning
-gcv --gradient-clip-val	0.0	The maximum allowed norm of the gradient (0.0 means no clipping)
-wd --weight-decay	0.0	The strength of adding L2-loss to the fine-tuning loss
-eas --eval-acc-steps	1	Number of accumulation samples during evaluation and testing
-md --model-dir	''	Directory name for the model to save checkpoints and testing results
-et --error-type	''	The error type for fine-tuning or testing
-stl --save-total-limit	-1	Maximum number of checkpoints to save
-pt --pre-trained	True	Whether to use the pre-training model or not

The tfix_testing.py has the following arguments: Required arguments:

 -lm --load-model   Path to the model's checkpoint that will be tested.

Optional arguments:

Parameter	Default	Description
-bs --batch-size	1	Batch size to be used in fine-tuning
-lm -–load-model	''	The path to a trained model to run testing
-ea --eval-all	False	Whether to evaluate model on random test or not
-eas --eval-acc-steps	1	Number of accumulation samples during evaluation and testing
-md --model-dir	''	Directory name for the model to save checkpoints and testing results
-et --error-type	''	The error type for fine-tuning or testing

To fit the program in memory, you should adjust arguments like -bs based on your machine configuration. Below you can find detailed information for some of the parameters.

Parameter -lm

The code treats the fine-tuning and testing as entirely separate procedures. This means that once the fine-tuning is done, the testing procedure will not start automatically. To test a trained model, you need to use this parameter and pass a valid checkpoint (saved during fine-tuning and can be found in the model's directory). When the -lm flag is not empty, the code will automatically switch to testing mode.

Parameter -ea

There are two testing dataset: clean test and random test. When the flag -ea is set to false, the testing runs only on clean test. If it is set to true, the testing is done on the large random test.

Parameter -ds

This parameter is used during testing and specifies the chunk size of the testing dataset. If this parameter is set, the testing dataset will be split into multiple chunks given the chunk size to fit the program in memory.

Parameter -eas

This parameter does not affect testing or fine-tuning but can save you from memory overflows. The testing is done on GPUs, and the samples will be pushed back to the CPU every -eas step to open up space in GPU memory. So you can try to lower this value if you encounter memory overflows during validation and testing.

Parameter -md

The model directory used for saving fine-tuning checkpoints or testing results. You can specify a directory name or a default directory name consists of date and time is given. It would help if you use this parameter to name the experiments that you run.

Parameter -et

This parameter specifies the error type when fine-tuning or testing per error type.

Parameter -pt

Either you can train T5 from scratch, or you can fine-tune it. Setting the flag -pt to true uses pre-trained model.

An example for testing

python tfix_testing.py -mn t5-large -lm data_and_models/models/t5large -md t5large_test

The testing results consist of two files first_accs.txt and test_data.json, and are saved in the model directory. first_accs.txt reports the exact match accuracy for each error type and test_data.json stores the TFix's output fixes in JSON format.

An example for fine-tuning

python tfix_training.py -e 30 -bs 32 -mn t5-large -md t5large_new

You can use the CUDA_VISIBLE_DEVICES flag to control the GPUs used for fine-tuning or testing.

Reproducing the experiment results

We provide scripts to obtain the exact match accuracy in our experiments. Obtaining the error removal accuracy involves other complex logics (e.g., calling ESLint) which we plan to release in the future.

Experiment: Model Size

Obtaining testing results with provided trained models:

python tfix_testing.py -mn t5-large -lm data_and_models/models/t5large -md t5large_test
python tfix_testing.py -mn t5-base -lm data_and_models/models/t5base -md t5base_test
python tfix_testing.py -mn t5-small -lm data_and_models/models/t5small -md t5small_test

You can also fine-tune the models by yourself:

python tfix_training.py -e 30 -bs 32 -mn t5-large -md t5large_new
python tfix_training.py -e 30 -bs 32 -mn t5-base -md t5base_new
python tfix_training.py -e 30 -bs 32 -mn t5-small -md t5small_new

Experiment: No Pre-training

Obtaining testing results with the provided trained model:

python tfix_testing.py -mn t5-large -lm data_and_models/models/t5large-no-pretrain -md t5large-no-pretrain_test

You can also train the model by yourself:

python tfix_training.py -e 30 -bs 32 -mn t5-large -md t5large-no-pretrain_new -pt False

Experiment: Fine-tuning per error type

We show how to perform the experiment for the guard-for-in error type. For other errors, simply change the -et argument. For disk space reasons, we only provide the fine-tuned model for the guard-for-in type.

Obtaining testing results with the provided trained model:

python tfix_testing.py -mn t5-large -lm data_and_models/models/t5large_guard-for-in -et guard-for-in -md t5large_guard-for-in_test

You can also fine-tune the model by yourself.

python tfix_training.py -e 30 -bs 32 -mn t5-large -md t5large_guard-for-in_new -et guard-for-in