Yesterday, GitHub introduced the Experiments platform for sharing demonstrations of their research projects and the idea behind them. With this platform, it aims to give the end users “insight into their research and inspire them to think audaciously about the future of software development”.
Why has GitHub introduced ‘Experiments’?
Just like Facebook and Google, GitHub regularly conducts research in machine learning, design, and infrastructure. The resultant products are rigorously evaluated for stability, performance, and security. If these products meet the success criteria for product release, they are then released for end users. Experiments will help GitHub share details about their research as they happen.
‘Semantic Code Search’: The first demo published on Experiments
The GitHub researchers also published their first demo of an experiment called Semantic Code Search. This system helps you search code on GitHub using natural language.
How does Semantic Code Search work?
The following diagram shows how Semantic Code Search works:
Step1: Learning representations of code
In this step, a sequence-to-sequence model is trained to summarize code by supplying (code, docstring) pairs. The docstring here is the target variable the model is trying to predict.
Step 2: Learning representations of text phrases
Along with learning representations of code, the researchers wanted to find a suitable representation for short phrases. To achieve this, they trained a neural language model by leveraging the fast.ai library. Using the concat pooling approach, the representations of phrases were extracted from the trained model by summarizing the hidden states.
Step 3: Mapping code representations to the same vector-space as text
In this step, the code representations learned from step 1 were mapped to the vector space of text. To accomplish this they fine-tuned the code-encoder.
Step 4: Creating a semantic search system
The last step is to bringing everything together to create a semantic search mechanism. The vectorized version of all code is stored in a database, and nearest neighbor lookups are performed to a vectorized search query.