Skip to content


Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

Automatic WAT Discovery

This is an ongoing project to automatically discover code behavior inconsistencies (WATs) between Python and R. This repo contains modules to parse Kaggle jupyter notebooks/scripts from competitions like Titanic in order to find similar code snippets across Python and R.


It's easiest to use baker which makes use of VirtualBox to quickly spin up a VM for the project. Once baker/VirtualBox is installed on your computer, simply execute the following in the root directory:

baker bake --local .

This sets up a VM for the project synced with the local folder and installs Python and R including all dependencies (requirements.txt).

Running with baker:

Then execute the following in the root directory:

baker run [cmd]

where cmd runs one of the phases described below and detailed in the default commands within baker.yml. For example, to filter all the scripts for Python snippets, one can execute:

baker run filterPy

Running modules directly:

You can also directly run the modules if you need more control or for testing.

Within root:

python [py|r]

python [kaggle|experiments] [-s single_dataframe | -r random_dataframes] [number of inputs to test <= 256]

where number of inputs are how many arguments in total you want to produce when using -r. If supplying -s, there is no need for supplying an additional number; a CSV template or dataframe template is used.

python SIM_T (<= 1.0) [keep]

where SIM_T is the similarity score and keep is an optional argument to also store test results and their scores.

To query from Python/R snippets with test cases and results, the cluster use the module:

python cluster_file low_score high_score edit_score

where cluster_file is the name of the file within the files directory (.csv is not required), low_score is the lower bound of the overal similarity score, the high_score is the upper bound and the edit_score is the syntactic distance between the snippets. For e.g., running:

python clusters_0.3 .6 .9 .5


df.iloc[0:5, 0:3] df[1:5, 1:3]
Row score: 0.6
Column score: 1.0
Overall score: 0.6
Edit distance: 0.794
Test case:

   col0  col1  col2  col3
0     0     6   NaN     5
1     1     3  ID_6     7
2     2     3  ID_4     7
Python output:

   col0  col1  col2
0     0     6   NaN
1     1     3  ID_6
2     2     3  ID_4
R output:

         col0        col1  col2
0           0           6   NaN
1           1           3  ID_6
2           2           3  ID_4
3 -2147483648 -2147483648   NaN
4 -2147483648 -2147483648   NaN

Within py_side:

python [notebook|script]

python [number of inputs to test <= 256] [all | dataframe | series | array]

where the 3rd argument specifies where you want to filter outputs; supply all to not filter any particular data type.

Within r_side:

python [notebook|script]

python [number of inputs to test <= 256] [all | dataframe | series | array]

Execution Phases:

The modules are run in order according to the following phases:

1. Preparation phase

First, the Kaggle Notebooks are traversed and the file paths are gathered for both Python and R. These lists will be used in the next phase. Relevant Files:

  • to traverse Notebooks and create file path lists for both Python/R Notebooks/Scripts stored in files

2. Segmentation + Filter + Normalization phase

Next, both the Python/R notebooks/scripts are segmented, where each line is considered a candidate expression. These candidates are then filtered for one-liner stand-alone expressions, discarding block expressions that span multiple lines like if or def in Python, or function and for in R. In this filtering process, each line must also fit a subset of the the grammar for each language (Python/pandas and R). Once the expressions meet these requirements, they are noramlized: 1) Dataframe variable names are renamed to df to standardize the dataframe variables for execution in the Execution phase 2) Whitespace within the expressions are stripped to unbias during the calculation of syntactical edit distances in the Cluster phase.

For Python, the built-in ast module is used to parse Python code and filter for certain expressions using the ast.Visitor class. The filtered expressions are then normalized using the ast.Transformer class. Relevant files:

  • filters and stores a csv file containing snippets in python_side
  • contains the Visitor/Transformer

For R, the rpy2 is used in Python code to use R's getParseData() function to filter for one-liner expressions and that meet a small subset of grammar. Then, a custom R script is used to normalize the expressions. Relevant files:

3. Input Generation + Execution phase

To execute Python/R snippets, inputs are generated which are dataframes based on a template csv file (for e.g. train.csv for the titanic competition). The generated dataframes are psuedo-random as column labels are preserved as well as column types. Only the values are either randomly generated within bounds or shuffled in the case of levels (for e.g. Sex as 0/1).

Values for int/float column types are randomly generated within the min/max values of the column; string values are randomly shuffled for str column types; and some NaN values are added if any NaN existed in the template's column. For both Python/R, the is used to generate arguments using pandas and numpy.

Then the Python/R snippets are executed against these generated input dataframes.

For Python, compile is used to compile an expression and eval to execute. Relevant files:

  • to generate, execute and store execution results into a py_dfs.pkl file in files

For R, rpy2.robjects.globalenv is used to introduce the argument into the embedded R's environment and rpy2.robjects.r is used to evaluate the expression. Relevant files:

  • to generate, execute and store execution results into a r_dfs.pkl file in files

4. Cluster phase

Finally, the Python/R snippets are then clustered according to output similarity score from 0 to 1. For scalars like ints/floats, a size_diff is calculated, for booleans the score is either 0 or 1, and for strings, the jaccard similarity score is calculated. For dataframes, the largest common area (LCA) dimension is first determined. Then the LCA is used as a window to slide the smaller dataframe over the larger to find the region with the highest cell similarity. Various different measures could be included for dataframes such as similarity by columns or rows. The edit distance between the Python and R snippets is also calculated using various measures like levenshtein or jaro. The clustered snippets are then stored in a csv file. Relevant files:

Notes on performance:

Most of the phases (2-4) can be time-consuming. Python's multiprocessing module is used to speed up the computations so that time execution is reduced considerably especially when running in a laptop.


Automatically identify code behavior differences across Python and R corpus from Kaggle.






No releases published


No packages published