Sampling k paths from AST tree about code2seq HOT 11 OPEN

Hi @Avra2 ,

Here it is:
https://github.com/tech-srl/code2seq/blob/master/reader.py#L78-L87

The code is not as simple as one would imagine, because every example has a different number of paths, so we need to sample the paths of every example separately.

Let me know if you have any questions!
Uri

from code2seq.

@urialon. Thank you.

You mentioned in your paper that, "we represent node using a learned embedding matrix E^nodes and then encode the entire sequence using the final states of a bi-directional LSTM." What is embedding matrix E^nodes and how you get that in your code please?

from code2seq.

@urialon. Hello Doctor,

Can you please just add some hints what algorithm you used to sample over AST paths? What format your AST path is in in the first place? Have you used any AST parser to generate the AST tree, then you looped over that parsed AST tree to generate paths? For binary tree, I see it's easy to loop over it to generate all paths between all terminals, but not sure how you did that given that AST is an N-tree.

Thanks.

from code2seq.

Hi @Avra2 ,
See answers below.

What is embedding matrix E^nodes and how you get that in your code please?

This is a standard, simple, embedding matrix, initialized here:
https://github.com/tech-srl/code2seq/blob/master/model.py#L359
and used here: https://github.com/tech-srl/code2seq/blob/master/model.py#L519-L520

Can you please just add some hints what algorithm you used to sample over AST paths?

A simple sampling without repetitions.

What format your AST path is in in the first place?

See the README and the processed datasets.

Have you used any AST parser to generate the AST tree, then you looped over that parsed AST tree to generate paths?

Yes, we used JavaParser in our code, but basically every parser can work, the parser is only used at preprocessing time. The neural network is agnostic to the parser.

For binary tree, I see it's easy to loop over it to generate all paths between all terminals, but not sure how you did that given that AST is an N-tree.

Why is it different? For every pair of terminals, we can go up in the tree until we reach their common ancestor, and that is the shortest path.

Let me know if you have any more questions
Uri

from code2seq.

Hi @urialon,
I am using your technique in my research work. I want to split path into Right path and left path. Kindly guide me, how can I do this. Thanks in anticipation.

from code2seq.

Hi @sadiasahar ,
Thank you for your interest in our work.

Can you please provide more details?
Uri

from code2seq.

@urialon, thank you for the reply.
I can explain it with an example. Suppose the path between node elements and true is:

(elements, Name↑ FieldAccess↑ Foreach↓ Block↓ IfStmt↓ Block↓ Return↓ BooleanExpr, true)

Now I want to split path from common ancestor. In this Example, from 'Foreach' Node. The Output I am requiring is

(elements, [Name↑ FieldAccess↑], [Foreach↓ Block↓ IfStmt↓ Block↓ Return↓ BooleanExpr], true)

from code2seq.

I see. We tried something similar in this PLDI 2018 paper.

So what's the problem? Can you just edit the data files and split the paths?

from code2seq.

When I preprocessed my data, I got the output

['i', 'Nm0|Ls|For|VDE|Prim0', 'int']

The problem is, in the string

'Nm0|Ls|For|VDE|Prim0'

I am unable to locate common ancestor, that is why can not split the path according to my requirement. How can I locate the ancestor in the string?

from code2seq.

Right, I see. You can change these lines:
https://github.com/tech-srl/code2seq/blob/master/JavaExtractor/JPredict/src/main/java/JavaExtractor/FeatureExtractor.java#L25-L26

But then you'll need to re-build the java package, and re-preprocess the data.

Best,
Uri

from code2seq.

@urialon, thank you so much. I tried as you suggested and got required output.

from code2seq.