Text2Code for Jupyter notebook

A proof-of-concept jupyter extension which converts english queries into relevant python code.

Blog post with more details:

Data analysis made easy: Text2Code for Jupyter notebook

Demo Video:

Text2Code for Jupyter notebook

Supported Operating Systems:

• Ubuntu
• macOS

Jupyter plugin Installation:

NOTE: We have renamed the plugin from mopp to jupyter-text2code. Uninstall mopp before installing new jupyter-text2code version.

pip uninstall mopp

GPU install

git clone https://github.com/deepklarity/jupyter-text2code.git
cd jupyter-text2code
pip install .

CPU-only install

For Mac and other Ubuntu installations not having a nvidia GPU, we need to explicitly set a environment variable at time of install.

git clone https://github.com/deepklarity/jupyter-text2code.git
export JUPYTER_TEXT2CODE_MODE="cpu"
cd jupyter-text2code
pip install .

Jupyter plugin Uninstallation:

pip uninstall jupyter-text2code

Usage Instructions:

• Open Jupyter notebook
• If installation happened successfully, then for the first time, Universal Sentence Encoder model will be downloaded from tensorflow_hub.
• Click on the Terminal Icon which appears on the menu (to activate the extension)
• Type "help" to see a list of currently supported commands in the repo
• Watch Demo video for some examples

Model training:

Generate training data:

From a list of templates present at jupyter_text2code/jupyter_text2code_serverextension/data/ner_templates.csv, generate training data by running the following command:

cd scripts && python generate_training_data.py

This command will generate data for intent matching and NER(Named Entity Recognition).

Create intent index faiss

Use the generated data to create a intent-matcher using faiss.

cd scripts && python create_intent_index.py

Train NER model

cd scripts && python train_spacy_ner.py

Steps to add more intents:

• Add more templates in ner_templates with a new intent_id
• Generate training data. Modify generate_training_data.py if different generation techniques are needed or if introducing a new entity.
• Train intent index
• Train NER model
• modify jupyter_text2code/jupyter_text2code_serverextension/__init__.py with new intent's condition and add actual code for the intent
• Reinstall plugin by running: pip install .

TODO:

• Publish Docker image
• Refactor code and make it mode modular, remove duplicate code, etc
• Add support for Windows
• Add support for more commands
• Improve intent detection and NER
• Explore sentence Paraphrasing to generate higher-quality training data
• Gather real-world variable names, library names as opposed to randomly generating them
• Try NER with a transformer-based model
• With enough data, train a language model to directly do English->code like GPT-3 does, instead of having separate stages in the pipeline
• Create a survey to collect linguistic data
• Add Speech2Code support

Authored By:

• Deepak Rawat
• Kartik Godawat