home/ atoms/ mcs-nn-latent-batch

mcs.nn~ packs all of one instance's inlets/outlets into a single multi-channel connection, enabling per-batch latent operations

mcs.nn~ is a multi-channel variant for batch operations on the latent dimensions of several sounds at once. Unlike mc.nn~ (which keeps separate inlets/outlets), mcs.nn~ packs every inlet/outlet of a single instance into one multi-channel connection, so the number of batches must be declared at initialization. Example: a decode function with 8 inputs — mcs.nn~ isis forward 1 has one inlet requiring 8 channels; to process 4 sounds at once use mcs.nn~ isis forward 4, with 4 inlets each carrying 8 channels. Performance is identical to mc.nn~; the point is that separated threads let you apply different operations to each batch — e.g. adding a different latent-noise vector per batch — which is tedious with mc.nn~.

Examples

mcs.nn~ isis forward 4 decodes 4 sounds of 8 latent dimensions each; add a different noise vector to each batch’s latent channels through the single mc wire.

Assessment

When would you choose mcs.nn~ over mc.nn~? Give a latent-manipulation task that is convenient in mcs.nn~ but tedious in mc.nn~, and state what must be declared at initialization.

“number of batches must be declared at initialization.”
corpus · neural-synthesis-in-max-8-with-rave-ircam-forum-tutorial · chunk 3