[Bug] Cannot feed value of shape - stereo/mono bug?

Hi!

I'm getting an error attempting to separate with a newly trained model:
Cannot feed value of shape (3113, 2049, 2) for Tensor 'mix_stft:0', which has shape '(None, 2049, 1)'

I'm guessing the third number is number of channels? Both the training and the file to separate were mono, so there shouldn't be a 2 anywhere. Last time I experimented with training it didn't work on mono; that seems to be fixed in the current build, but possibly there's still a hardcoded 2 somewhere in the separate code?

I've also tried it on audio_example.mp3 and audio_example_mono.mp3 with the same result - only the first number changes, which seems to be the duration of the file?

Reproduce separation with audio_example.mp3 and audio_example_mono.mp3, then trace the separate path around the reported mix_stft:0 shape mismatch. Compare the channel dimensions produced for mono input with the trained model's expected shape; done means both example files separate without the (3113, 2049, 2) versus (None, 2049, 1) error.

Written by the indexing model from the issue text.