Tobias Pitters

@connortann: I would like to ask for your advice. So this is a pretty longstanding issue, and currently I am capable of evaluating shap values for LSTM, GRU and SimpleRNN...

The errors seem mainly to happen in 2 tests and whether the shap values sum up as expected is heavily dependent on the input. Not sure if I should mark...

> Hi @CloseChoice @connortann I am very closely following this issue and I am happy to test the releases or fixes done Kindly let me know if you need any...

> Hi @CloseChoice > > Sure, Happy to do so. > > Just double checking I am good If I run this setup: > > this experiment --> https://www.kaggle.com/code/billlucas/explaining-cnn-lstm-using-shap/notebook With...

> > hi @CloseChoice > > i am using it for some LSTM code, as i couldn't compute SHAP values. Now code is working, but I am getting the gradients...

> What about call to `self._init_between_tensors(out.op, shap_rAnD)` in grad_graph function? I changed it to `self._init_between_tensors_eager(out.op, shap_rAnD, data..?)` but then got other errors. maybe if that is fixed, gradients would be...

> @CloseChoice Do we have a status update on this? Would really appreciate it because my model is using some tf2 features and require shap lstm explainer. Hey, thanks for...

Note: take a deep look how this repo does it: https://github.com/alewarne/Layerwise-Relevance-Propagation-for-LSTMs

close this for now, there is another approach that we need to take here, hopefully I'll ever get to it ;)

I am looking into this PR right now. It's a bit hard to actually see the difference/similarities to the existing implementation, will try to change a few things there. Could...