gaozhihan/PreDiff

Processing of four-dimensional data

qwertyui20 opened this issue · 1 comments

Can four-dimensional data be used for prediction on Prediff?

Thank you for your question. Are you referring to data tensors with the shape (batch_size, time, height, width, altitude, channels), where altitude represents the axis perpendicular to the ground plane? Unfortunately, the current design of PreDiff does not support processing data in this specific format. However, a simple workaround is to merge the altitude axis into the channels axis. This can be done by reshaping the tensor from b t h w a c to b t h w (a c). This approach is often feasible since the dimension of the altitude axis is typically much smaller than the height and width dimensions.