Compute saliency map with an onnx model

[MichaelBetser]

Hi all,

I was wondering if it was possible to compute a forward and backward pass using onnx runtime/onnx models. This is for the purpose of explainability: in the end I would like to output both the prediction and the saliency map (gradients * input embeddings).

Thanks,

Michael.