Python Autocomplete

This project try autocompleting python source code using LSTM or Transformer models.

Training model:

Evaluating trained model:

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.

The dataset we use is the python code found in repos linked in Awesome-pytorch-list. We download all the repositories as zip files, extract them, remove non python files and split them randomly to build training and validation datasets.

We train a character level model without any tokenization of the source code, since it's the simplest.

1. Clone this repo
2. Install requirements from requirements.txt
3. Run python_autocomplete/create_dataset.py.
   • It collects repos mentioned in PyTorch awesome list
   • Downloads the zip files of the repos
   • Extract the zips
   • Remove non python files
   • Collect all python code to data/train.py and, data/eval.py
4. Run python_autocomplete/train.py to train the model. Try changing hyper-parameters like model dimensions and number of layers.
5. Run evaluate.py to evaluate the model.

You can also run the training notebook on Google Colab.

Here's a sample evaluation of a trained transformer model.

Colors:

• yellow: the token predicted is wrong and the user needs to type that character.
• blue: the token predicted is correct and the user selects it with a special key press, such as TAB or ENTER.
• green: autocompleted characters based on the prediction

We are working on a simple extension for VSCode for demonstration.