Quickstart

1. First we will create a virtual environment with anaconda.
   If you do not have anaconda installed, please check the installation instructions here.

      $ conda create --name probe_env python=3.8
   

2. Next we will activate the virtual environment.

      $ conda activate probe_env
   

3. Once we are done with that, we will clone this github repository and cd into it.

      $ git clone https://github.com/giganticode/inspect.git && cd inspect/
   

4. Next we will install all the necessary packages in the requirements.txt file

      $ pip3 install -r requirements.txt
   

5. We will then begin generating input feature vectors from the pre-trained model weights.

      $ CUDA_VISIBLE_DEVICES=0 python3 probe_extractor.py
   

6. Finally, we will use the pre-trained vectors to make some predictions.

      $ CUDA_VISIBLE_DEVICES=0 python3 probe_classifier.py > results.txt
   

Hurray! You're done! 🎉
The final task accuracies of each model (layer by layer) will be saved in results.txt file.

NOTE: In steps 5 & 6, we assume that you have access to a gpu with the cuda toolkit installed. If you are using a cpu instead, you can simply run the python scripts without the CUDA_VISIBLE_DEVICES=0 part.

---

Model Performance by Tasks

Code Length Prediction

AST Node Tagging Task

Cyclomatic Complexity

Invalid Type Prediction

Overall Results via Heatmaps

---

Acknowledgements

This repository uses and is built upon the following works:

• pytorch-pretrained-BERT from huggingface
• SentEval from facebookresearch
• bert-syntax from yoavg
• tpdn from tommccoy1
• rnn_agreement from TalLinzen
• Chu-Liu-Edmonds from bastings
• interpret_bert from ganeshjawahar

License