search

ajsanjoaquin / Shapley_Valuation

PyTorch reimplementation of computing Shapley values via Truncated Monte Carlo sampling from "What is your data worth? Equitable Valuation of Data" by Amirata Ghorbani and James Zou [ICML 2019]
MIT License
25 stars 6 forks source link

Model training is missing #1

Open eardic opened 2 years ago

eardic commented 2 years ago

Why the model is not trained with the new batch at the line below?

https://github.com/ajsanjoaquin/Shapley_Valuation/blob/bf8cf85d3f8a288fdb43ca1c996be18dac25c56b/tmc.py#L154

ajsanjoaquin commented 2 years ago

Hello @eardic, thanks for spotting the bug. I agree that it should train the model before computing the new score.

Looking at the paper (https://arxiv.org/pdf/1904.02868.pdf), page 4, it says to add a new point in our subset and see if performance (e.g. accuracy) improves compared to the performance with the previous subset.

I'm tied with other projects for now, so I would appreciate if you can submit a PR to fix this.