PLUR (Programming-Language Understanding and Repair) is a collection of
source code datasets suitable for graph-based machine learning. We provide
scripts for downloading, processing, and loading the datasets. This is done
by offering a unified API and data structures for all datasets.


mkdir -p ${SRC_DIR} && cd ${SRC_DIR}
# For Cubert.
git clone --depth=1
export PYTHONPATH=${PYTHONPATH}:${SRC_DIR}/google-research
git clone && cd plur
python -m pip install -r requirements.txt
python install

Test execution on small dataset

cd plur
python3 --dataset_name=manysstubs4j_dataset \
  --stage_1_dir=/tmp/manysstubs4j_dataset/stage_1 \
  --stage_2_dir=/tmp/manysstubs4j_dataset/stage_2 \
  --train_data_percentage=40 \
  --validation_data_percentage=30 \


Basic usage

Data generation (step 1)

Data generation is done by calling plur.plur_data_generation.create_dataset().
The data generation runs in two stages:

  1. Convert raw data to plur.utils.GraphToOutputExample.
  2. Convert plur.utils.GraphToOutputExample to TFExample.

Stage 1 is unique for each dataset, but stage 2 is the same for almost all datasets.

from plur.plur_data_generation import create_dataset

dataset_name = 'code2seq_dataset'
dataset_stage_1_directory = '/tmp/code2seq_dataset/stage_1'
stage_1_kwargs = dict()
dataset_stage_2_directory = '/tmp/code2seq_dataset/stage_2'
stage_2_kwargs = dict()
create_dataset(dataset_name, dataset_stage_1_directory, dataset_stage_2_directory, stage_1_kwargs, stage_2_kwargs) also provides a command line interface, but it offers less flexibility.

python3 --stage_1_dir=/tmp/code2seq_dataset/stage_1 --stage_2_dir=/tmp/code2seq_dataset/stage_2

Data loader (step 2)

After the data is generated, you can use PlurDataLoader to load the data. The data loader loads TFExamples but returns them as numpy arrays.

from plur.plur_data_loader import PlurDataLoader
from plur.util import constants

dataset_stage_2_directory = '/tmp/code2seq_dataset/stage_2'
split = constants.TRAIN_SPLIT_NAME
batch_size = 32
repeat_count = -1
drop_remainder = True
train_data_generator = PlurDataLoader(dataset_stage_2_directory, split, batch_size, repeat_count, drop_remainder)

for batch_data in train_data_generator:
  # your training loop...

Training (step 3)

This is the part where you use your own model to train on the PLUR data.

The models and the training code from the PLUR paper are not yet part of the
current release. We plan to release it in the near future.

Evaluating (step 4)

Once the training is finished, you can generate the predictions on the test data and use to evaluate the performance. works in offline mode, meaning that it expects a file containing the ground truths, and a file containing the predictions.

python3 --dataset_name=code2seq_dataset --target_file=/tmp/code2seq_dataset/targets.txt --prediction_file=/tmp/code2seq_dataset/predictions.txt

Transforming and filtering data

If there is something fundamental you want to change in the dataset, you should apply them in stage 1 of data generation, otherwise apply them in stage 2. The idea is that stage 1 should only be run once per dataset (to create the plur.utils.GraphToOutputExample), and stage 2 should be run each time you want to train on different data (to create the TFRecords).

All transformation and filtering functions are applied on plur.utils.GraphToOutputExample, see plur.utils.GraphToOutputExample for more information.

E.g. a transformation that can be run in stage 1 is that your model expects that graphs in the dataset have no loop, and you write your transformation function to remove loops. This will ensure that stage 2 will read data where the graph has no loops.

E.g. of filters that can be run in stage 2 is that you want to check your model performance on different graph sizes in terms of number of nodes. You write your own filter function to filter graphs with a large number of nodes.

from plur.plur_data_generation import create_dataset

dataset_name = 'code2seq_dataset'
dataset_stage_1_directory = '/tmp/code2seq_dataset/stage_1'
stage_1_kwargs = dict()
dataset_stage_2_directory = '/tmp/code2seq_dataset/stage_2'
def _filter_graph_size(graph_to_output_example, graph_size=1024):
  return len(graph_to_output_example.get_nodes()) <= graph_size
stage_2_kwargs = dict(
create_dataset(dataset_name, dataset_stage_1_directory, dataset_stage_2_directory, stage_1_kwargs, stage_2_kwargs)

Advanced usage

plur.plur_data_generation.create_dataset() is just a thin wrapper around plur.stage_1.plur_dataset and plur.stage_2.graph_to_output_example_to_tfexample.

from plur.plur_data_generation import create_dataset

dataset_name = 'code2seq_dataset'
dataset_stage_1_directory = '/tmp/code2seq_dataset/stage_1'
stage_1_kwargs = dict()
dataset_stage_2_directory = '/tmp/code2seq_dataset/stage_2'
stage_2_kwargs = dict()
create_dataset(dataset_name, dataset_stage_1_directory, dataset_stage_2_directory, stage_1_kwargs, stage_2_kwargs)

is equivalent to

from plur.stage_1.code2seq_dataset import Code2seqDataset
from plur.stage_2.graph_to_output_example_to_tfexample import GraphToOutputExampleToTfexample

dataset_name = 'code2seq_dataset'
dataset_stage_1_directory = '/tmp/code2seq_dataset/stage_1'
dataset_stage_2_directory = '/tmp/code2seq_dataset/stage_2'
dataset = Code2seqDataset(dataset_stage_1_directory)

dataset = GraphToOutputExampleToTfexample(dataset_stage_1_directory, dataset_stage_2_directory, dataset_name)

You can check out plur.stage_1.code2seq_dataset for arguments relevant for code2seq dataset. For example code2seq dataset provides java-small, java-med and java-large datasets. Therefore you can create a java-large dataset in this way.

from plur.stage_1.code2seq_dataset import Code2seqDataset

dataset_name = 'code2seq_dataset'
dataset_stage_1_directory = '/tmp/code2seq_dataset/stage_1'

dataset = Code2seqDataset(dataset_stage_1_directory, dataset_size='large')

Adding a new dataset

All datasets should inherit plur.stage_1.plur_dataset.PlurDataset, and placed under plur/stage_1/, which requires you to implement:

  • download_dataset(): Code to download the dataset, we provide download_dataset_using_git() to download from git and download_dataset_using_requests() to download from a URL, which also works with a Google Drive URL. In download_dataset_using_git() we download the dataset from a specific commit id. In download_dataset_using_requests() we check the sha1sum for the downloaded files. This is to ensure that the same version of PLUR downloads the same raw data.
  • get_all_raw_data_paths(): It should return a list of paths, where each path is a file containing the raw data in the datasets.
  • raw_data_paths_to_raw_data_do_fn(): It should return a beam.DoFn class that overrides process(). The process() should tell beam how to open the files returned by get_all_raw_data_paths(). It is also here we define if the data belongs to any split (train/validation/test).
  • raw_data_to_graph_to_output_example(): This function transforms raw data from raw_data_paths_to_raw_data_do_fn() to GraphToOutputExample.

Then add/change the following lines in plur/

from plur.stage_1.foo_dataset import FooDataset

flags.DEFINE_enum('dataset_name', 'dummy_dataset',
                  ['code2seq_dataset', 'dummy_dataset',
                   'funcom_dataset', 'great_var_misuse_dataset',
                   'manysstubs4j_dataset', 'foo_dataset'],
                  'Name of the dataset to generate data.')

def get_dataset_class(dataset_name):
  """Get the dataset class based on dataset_name."""
  if dataset_name == 'code2seq_dataset':
    return Code2SeqDataset
  elif dataset_name == 'dummy_dataset':
    return DummyDataset
  elif dataset_name == 'funcom_dataset':
    return FuncomDataset
  elif dataset_name == 'great_var_misuse_dataset':
    return GreatVarMisuseDataset
  elif dataset_name == 'hoppity_single_ast_diff_dataset':
    return HoppitySingleAstDiffDataset
  elif dataset_name == 'manysstubs4j_dataset':
    return ManySStuBs4JDataset
  elif dataset_name == 'foo_dataset':
    return FooDataset
    raise ValueError('{} is not supported.'.format(dataset_name))

Evaluation details

The details of how evaluation is performed are in plur/eval/


Licensed under the Apache 2.0 License.


This is not an officially supported Google product.


Please cite the PLUR paper, Chen et al.


View Github