Similar to the top answer at: How to control if input features contribute exclusively to one neuron in subsequent layer of a Tensorflow neural network?
, I know that I can throw each single categorical variable as its own Input Layer and then concatenate after the embedding has been done before channeling the input further into the model.
However, given the nature of auto encoders, I then also have to reconstruct every input and calculate the reconstruction loss for each.
Instead of a multi-input network, assuming I have 10 categorical variables, is there a way I can split a single input into 10 embedding layers while making sure each categorical variable gets isolated in its own embedding layer (thereby keeping the separate nature of having 1 input for each variable instead of mixing them) - Thereby resulting in just 1 total reconstruction of the whole input.
