Hello, I’m using a variational autoencoder in a project and I wanted to ask some opinions about few issues:
My inputs are made by the concatenation of two sparse vectors which (independently) sum to one (So they belong to some $D$-dimensional simplex).
I’m not sure about which loss function might be suitable to reconstruct these kind of data, I’ve actually read that
torch.nn.CosineLossEmbedding()could be a good choice but I’m not seeing noticeable results for now.
I’ve tried to eliminate biases from all my linear layers.
As my input consists of two separate vectors, could it make sense to provide a stack of different encoders/decoders to the architecture (2 in this case) to be sure that they are somehow processed separately?