How to jointly embed multi-dimensional discrete features?


I am trying to use embedding layers to learn continuous representations for discrete input features. Each of these discrete features is multidimensional. That is, one sample in the input batch contains two integer items for example. Is there a possibility to use one embedding layer whose continuous representation dictionary can be accessed using the two integer values?

I have tried to use two separated embedding layers. But, unfortunately, I get an issue of “RuntimeError: CUDA error: device-side assert triggered”