Closed alanngnet closed 1 month ago
Impact is significant for large datasets. Throughput rapidly decreases after a few thousand samples, as dealing with overflowing memory takes more and more resources. Eventually I have to abort the run and restart the extract_csi_features script.
Added workaround to just do CQT processing in batches and then run garbage collection. Batch size=5000 worked well on my system. It does noticeably bog down before 5,000, but running gc more often also adds significant delay even just to run through lines that don't need new CQT generation.
Something is causing gradually increasing memory consumption for each python instance running during CQT generation, at least in MPS context. Possibly a bug in nnAudio to report those authors, or already a known bug? Or something we can address in our project?