woven spire
Hi, did anyone try ADTOF? I got it running, but it's at 0.2x realtime on my RX 9070 XT, which feels really weird. Or is it just this slow?
https://github.com/xavriley/ADTOF-pytorch
The normal one at https://github.com/MZehren/ADTOF didn't run because tf-keras was incompatible with the models I think.
crystal flower
Hi
What versions of OS and pytorch do you use?
woven spire
I'm on EndeavourOS and kernel 6.18.4-arch1-1, Pytorch 2.11.0a0+ROCm 7.11.0a20260107
I didn't yet update ROCm, I have a potentially fragile venv I'm developing in.
System is Ryzen 9 9900X with 96GB of 6000MHz RAM.
woven spire
Here is what Claude gave me:
https://github.com/Taikakim/mir-feature-extraction/blob/main/src/transcription/drums/adtof.py
I even modified ADTOF itself to apparently a faster FFT implementation, but total processing time is still slow.
So, it's entirely possible I'm doing something weird too, since I'm not a coder.
crystal flower
I just tested ADTOF-pytorch on W7900 and speed is 3.5x realtime. Could you update ROCm+pytorch and run adtof again. During script execution check GPU utilization via rocm-smi