[Discussion] Exporting a checkpoint with static-sized tensors

I'm trying to convert the 4stems checkpoint to TensorFlow-lite and I need the model to have static-sized tensors (so as to be able to use the Metal or CoreML delegates).
I've tried specifying the input shape of 'waveform' but the outputs are still dynamic-sized.

Can this be done with the current checkpoints?

Start with the 4stems checkpoint conversion and the waveform input-shape setting described in the report. Check whether TensorFlow Lite export preserves static output shapes for the current checkpoints and record a reproducible result or limitation.

Written by the indexing model from the issue text.