dmitrykazhdan / representing-programs-with-graphs Goto Github PK

Looking at the Vocabulary creation code here:

• You need to get the most frequent (sub) tokens. By converting them into a set you lose count information. Use Counter instead and pass that into the Vocabulary.
• The current max_size is tiny. (a) Pull it out as a hyperparameter you can change (b) For "real" data, this should be larger. It's reasonable for fast, dry runs.

The placeholder node_token_ids seems off. The input should probably be of size batch_size x max_num_subtokens instead of batch_size x hidden_size. Essentially, the placeholder should contain the ids of the subtokens within the vocabulary, not their embeddings (you seem to be correctly looking up the embeddings of those ids here whose output should be batch_size x max_num_subtokens x hidden_size).

(we also discussed in person about masking, if I recall. The mask would be a 0/1 matrix of size batch_size x max_num_subtokens if you decide to implement this at some point)

Hello @dmitrykazhdan ,

I am a Master student at the HTWG Konstanz University of Applied Sciences in Computer Science.

Currently, I am working on my master thesis, where I am investigating various machine learning methods that deal with the naming of variables and methods.

My idea is to collect the found implementations within a proof of concept and to use them for a specific use case within the development process.

For example in pull request's to support the reviewers.
I want to publish the proof of concept as an open source project on GitHub.

Therefore my question:
Could you add a license to your repo?
May I use your implementation in my proof of concept?

Greetings
Marcel