Preserving the Characteristics of A Custom Graph

[akibzaman]

PLUR framework can accept a custom graph and produce the output task accordingly. Also, the framework is capable of incorporating relational information that can be used to represent syntax, data flow and control flow.

I had the following query regarding this point while exploring the PLUR framework:

If a graph with various relational edges are given as an input, does PLUR preserve the characteristics of the edges while feeding into the model or it creates a common outlook for all types of graph? In Summary, Does PLUR preserve the characteristics of a custom graph?

[dan-zheng]

Hi Akib,

I'm not an author of PLUR, but I'm happy to shed some light and am curious to understand your question better.

If a graph with various relational edges are given as an input, does PLUR preserve the characteristics of the edges while feeding into the model or it creates a common outlook for all types of graph? In Summary, Does PLUR preserve the characteristics of a custom graph?

I wonder what you mean by "characteristics of a custom graph"? What is custom about the graphs you have in mind – for example, do they have edges of different types? One example could be program graphs, where there are multiple edge types like NextLexicalUse or ComputedFrom.

Backing up a bit – as you may know, PLUR is not a particular model architecture, but rather a unified framework for code-related tasks, code representations, and model architectures. Thanks to the unification, PLUR makes it easy to reproduce results from various ML4Code papers (that each use different tasks, code representations, and models)¹, and also enables easy experimentation with new tasks, code representations, and models.

The PLUR paper shows experimental results for various model architectures, including edge-type-aware graph neural networks, in case that is what you mean by "custom graph characteristics". Namely, section 3.1 explains modeling choices: the GREAT model is implemented within PLUR as a graph encoder, and GREAT is a "relational Transformer" that biases a traditional Transformer architecture with relational information from graph edge types. Other edge-aware model types can be implemented within PLUR too, like GGNNs.

Does this answer your question?

Footnotes

1. Table 1 from the PLUR paper shows results from reproducing a wide range of ML4Code papers within the PLUR framework, achieving performance competitive with the original papers. ↩

[smoitra87]

Thanks @dan-zheng for the detailed explanation.

To add to the reply, yes the PLUR data loader and the modeling code does preserve edges and edge types when training the model. See the following relevant sections of code:

(1) Edges in GREAT var misuse
https://github.com/google-research/plur/blob/main/plur/stage_1/great_var_misuse_dataset.py#L150
(2) The data loader picks up the edge types from the vocab file
https://github.com/google-research/plur/blob/main/plur/plur_data_loader.py#L96
(3) Edges are read in and converted to tf.SparseTensor to be fed into the model.
https://github.com/google-research/plur/blob/main/plur/plur_data_loader.py#L307
(4) The model instantiates itself based on the maximum number of edge types.
https://github.com/google-research/plur/blob/main/plur/model_design/encoders.py#L64

Hope this helps.

[akibzaman]

Thanks a lot for the insightful answers @dan-zheng and for pointing out the code segments @smoitra87. It clarifies my query.