Hi,
Let’s assume that we have a generative model (e.g. GAN) such that it takes 1D input noise of length 1000 (i.e. vector of length 1000) and we need to generate a signal of the same length based on 1D condition vector of length 50.
My question: what is the best way to inject these 50 parameters into my model as a conditioning ?The most prominent technique I have found is to concatenate both the conditioning with the noise … but how this should be done in case of different lengths like this case ? Because the trivial concatenation in the channel dimension is not possible here … I have tried zero insertion to make the two vectors with the same length at the sequence dimension (noise and conditioning). But this is not effective.
Does it make sense to concatenate at the sequence dimension (i.e. to have one vector with length 150 and feed it to the network) ? Or there is no way in this case other than adding new layers for the conditioning data then concatenate the output tensors at the channel dimension ?
Best