Deeplift with embedding layers about deeplift HOT 2 CLOSED

Hi Greg,

I'd suggest two options:

(1) The advantage of using the DeepLIFT implementation in this repo is that it supports the RevealCancel rule, which we have found to be beneficial in some situations. However, the downside is that not all layers are supported, as you have discovered. If you are not attached to the RevealCancel rule, you can use the DeepLIFT implementation in the DeepExplain repository, or the DeepSHAP implementation in the shap repository (DeepSHAP is an extension of DeepLIFT to deal with several operations that DeepLIFT didn't support, but it doesn't have the RevealCancel rule)

(2) If you do want to use the RevealCancel rule, and all the other layers in your architecture are supported, you can bypass the embedding layer by structuring your network to take the embeddings themselves as inputs (i.e. you have a preprocessing step where you convert your original inputs to embeddings, and then supply the embeddings as the input to the network)

from deeplift.

Thanks for the quick reply. I was hoping to use the RevealCancel rule. Doing some preprocessing shouldn't be that big of a deal so I'll try that.

Thanks!

from deeplift.