Im trying to use a transformer to predict a sequence of X,Y coordinates. My question is how I would go about embedding the coordinates to a D-dimensional space in order to input it in my Transformer model.
So far I’ve done something like the following:
First I scale the coordinates to numbers between 0-100.
input_dim = 100 embedding_dim = 256 embedding = nn.Embedding(input_dim, embedding_dim) input_to_embed = torch.tensor([[20,30], [30, 20]]) ## [[X,Y],[X,Y]] two positions of x,y coordinates embed = embedding(input_to_embed) print(embed.shape)
However, when the coordinates have decimals (like [20.3, 30.4]), I get an error because the values does not have scalar type Long.
I suspect that this is not the best way to solve the problem so I am happy to hear your thoughts and suggestions!