LarsNet, converted to ONNX

ONNX conversion of the pretrained LarsNet drum-part separation networks (kick, snare, toms, hi-hat, cymbals), made for the open-source Noodle audio workstation.

Not for commercial use. These weights are licensed CC BY-NC 4.0, the same licence as the originals.

Credit

LarsNet is by Alessandro Ilic Mezza, Riccardo Giampiccolo, Alberto Bernardini and Augusto Sarti, Image and Sound Processing Lab, Politecnico di Milano.

What was changed

Only the format. Each of the five pretrained_*_unet.pth checkpoints was loaded into the authors' UNet (unet.py, unmodified) and its produce_mask method exported with PyTorch (opset 17, eval mode, fp32), without the optimizer state that the checkpoints also hold. The ONNX output matches PyTorch to about 1e-5. The weights are unchanged.

Interface (all five files)

  • input mag: float32 [1, 2, 2048, 512], the magnitude STFT of the stereo drum mix (n_fft 4096, periodic Hann, hop 1024, centred), 2048 bins (DC up to, not including, Nyquist), 512 frames (about 11.9 s at 44.1 kHz).
  • output mask: float32 [1, 2, 2048, 512], a gain in [0, 1]. Multiply it with the mix's spectrogram (the Nyquist bin gets a mask of 0), then inverse STFT with the mix's phase. The authors optionally apply a Wiener-style combination across the five masks.
file bytes sha256
kick.onnx 39346035 e7c994e9f191b7d66490bbc98561d368c8b86fcac8994582596dc028178c1fcf
snare.onnx 39346035 89d28d42d83316fe9da367a95b21538ddb5a8fdab2382b57ecfe6cc1f94a5b27
toms.onnx 39346035 3122013fdb5e303a97ec0f9c9b650b25670a8aab90a2877b3a50f9a4f34a9f6a
hihat.onnx 39346035 cacb88ee7f58d057d0c050972ab2a230450571a7f61c326aa5a79581327e42dd
cymbals.onnx 39346035 2eb483ea3eebb8f64d96a1a87fd00e2a537fefd699533308f9e398eef0b80173
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support