The developers behind TabNine have introduced a new deep learning model, “Deep TabNine,” which significantly improves suggestion quality. TabNine is a code autocompleter tool that predicts the rest of the code programmers are typing — much like word or phrase autocomplete in a Google search window. TabNine is a great time-saver that supports 23 mainstream programming languages and five code editors. News of the Deep TabNine addition went viral on social media and received over 4000 Twitter likes in just 24 hours.
Trained on about two million files on GitHub, the Deep TabNine model learns complex behaviors such as type inference in dynamically typed languages, and predicts tokens.
Deep TabNine also makes good use of trivial details that are difficult to obtain using traditional code completion tools. DeepTabNine can infer function names, parameters, and return types from documents written in natural language.
Additionally, Deep TabNine has incorporated a useful new feature that many users wanted: the ability to apply pre-existing knowledge to a small project or when a new library is being added to an existing project.
Deep TabNine is built upon GPT-2, a large transformer-based language model that can generate realistic paragraphs of text. The model, developed by San Francisco-based research company OpenAI, demonstrates compelling performance across a range of language tasks such as machine translation, question answering, reading comprehension and summarization.
Unlike other code completion plug-ins, Deep TabNine is automatically compiled based on a programmer’s past usage and habits, and includes the probability of the different predictions it provides. If a similar code appears in previous projects, Deep TabNine will also display the address directly in the completion candidate box.
A license is required to use TabNine for projects over 400kb and costs US$49 for individual users and US$99 for business users. Since the high-performance autocompleter is backed by a machine learning algorithm, it requires more than 10 billion FLOPS to run and cannot deliver a low latency experience for example on laptops. The high compute requirements prompted software developers to launch a TabNine Cloud server accelerated by cloud GPUs.
TabNine was developed by senior computer science student Jacob Jackson from the University of Waterloo, who received an IOI gold medal in 2014 and 2015. He posted on Reddit that TabNine is the first commercial software product he developed at the university.
The TabNine website with install instructions and other information is here.
Author: Reina Qi Wan | Editor: Michael Sarazen