Open lminer opened 2 years ago
Could you please provide some public models which are impacted now?
Most, current audio source separation models are impacted by this limitation as they use an audio signal representation called "complex as channels", where the real and imaginary parts of the complex spectrogram are extracted and stacked along the channels dimension. Some examples are: Demucs MDXNet LaSaft DCUNet
It would be great to get support for Real, Imag, and Complex. Anybody working with audio data has to deal with complex inputs and this is very difficult without support for these ops.