Neural Network Diagnostics

Debugging tools for pytorch neural networks

Fig.1 - Image classification (Resnet) diagnostics using integrated gradients and input gradients, model prediction label is "tiger cat".

Fig.2 - Language modeling (GPT2) diagnostics using integtated gradients, the input text is "Can humans dream of electric sheep ? This is a", next token prediction is "question".

Fig.3 - Language modeling (GPT2) diagnostics using input gradients, with same input and prediction.

Methods:

• Input Gradients, analog of the model coefficients for a deep network.
• Integrated Gradients, "Axiomatic Attribution for Deep Networks", Mukund Sundararajan Ankur Taly, Qiqi Yan [paper] [original implementation].

Demos

• ConvNets
• GPT2 LM

Environment setup

• Anaconda, for python package management.

    conda env create -f environment.yaml
    conda activate nn-diagnostics

• direnv, optional, used for automatic setup environment when changing directory.

Contribute

• Install pre-commit hooks

    pre-commit install