Skip to content
A simple neural network for python autocompletion
Branch: master
Clone or download

Latest commit

Latest commit 6e5649a May 13, 2020


Type Name Latest commit message Commit time
Failed to load latest commit information.
logs/simple_lstm/2a86d636936d11eab8740dffb016e7b1 🎉 initial commit May 12, 2020
parser 🎉 initial commit May 12, 2020
.gitignore 🎉 initial commit May 12, 2020
.labml.yaml 🎉 initial commit May 12, 2020
LICENSE 🎉 initial commit May 12, 2020 🎉 initial commit May 12, 2020 🎉 initial commit May 12, 2020
python-autocomplete.png 🎉 initial commit May 12, 2020 🎉 initial commit May 12, 2020
requirements.txt 🎉 initial commit May 12, 2020 ♻️ update lab May 13, 2020

This a toy project we started to see how well a simple LSTM model can autocomplete python code.

It gives quite decent results by saving above 30% key strokes in most files, and close to 50% in some. We calculated key strokes saved by making a single (best) prediction and selecting it with a single key.

We do a beam search to find predictions, upto ~10 characters ahead. So far it's too inefficient, if you are wondering about editor integration.

We train and predict on after cleaning comments, strings and blank lines in python code. The model is trained after tokenizing python code. It seems more efficient than character level prediction with byte-pair encoding.

A saved model is included in this repo. It is trained on tensorflow/models.

Here's a sample evaluation on a source file from validation set. Red characters are when a auto-completion started; i.e. user presses TAB to select the completion. The green character and and the following characters highlighted in gray are auto-completed. As you can see, it starts and ends completions arbitrarily. That is a suggestion could be 'tensorfl' and not the complete identifier 'tensorflow' which can be a little annoying in a real usage scenario. We can limit them to finish on end of tokens to fix that. Also you can notice that it completes across operators as well. Increasing the length of the beam search will let it complete longer pieces of code.

Try it yourself

  1. Clone this repo

  2. Install requirements from requirements.txt

  3. Copy data to ./data/source

  4. Run to collect all python files, encode and merge them into

  5. Run to evaluate the model. I have included a checkpoint in the repo.

  6. Run to train the model

You can’t perform that action at this time.