CQT frames are currently processed all in parallel but when dealing with big batches (or long audio files) this can generate a CUDA OOM. In this case, inputs should be split into chunks to automatically optimize parallelism while preventing OOM.
On a 1080, the limit is of about 9 minutes in parallel for a 16 kHz audio signal
CQT frames are currently processed all in parallel but when dealing with big batches (or long audio files) this can generate a CUDA OOM. In this case, inputs should be split into chunks to automatically optimize parallelism while preventing OOM.
On a 1080, the limit is of about 9 minutes in parallel for a 16 kHz audio signal