Tags: fletcherist/ddsp
Tags
Simplify and refactor RnnFcDecoder. Takes a list of input keys (from …
…the preprocessor output) and builds a fully connected stack for each of them.
Bump to version v0.2.0 as this will require old models to add a single line to their operative gin configs.
`RnnFcDecoder.input_keys = ('f0_scaled', 'ld_scaled')`
PiperOrigin-RevId: 309095767
Refactor losses accounting within model. * Remove use of `self.add_loss()` as it requires cumbersome keeping tracking of a parallel list of loss_names for tensorboard summaries. * Forward `__call__()` on model now has `return_losses` keyword argument that if True also returns a dictionary of all the losses on that pass, including 'total_loss'. * Needed to use the functional form (instead of a self.losses_dict property) to handle some tf.function() headaches that arose from side-effects. * Bumps version to v0.1.0. The version should have actually been bumped at v0.0.7 which made a minor revision change. PiperOrigin-RevId: 308137442