windowed-roformer-onnx

ONNX conversions of smulelabs/windowed-roformer (checkpoint mbr-win10-sink8; code and weights are MIT) for vocals/BGM separation.

Files

File	What it is
`windowed-roformer.onnx`	Plain fp32 conversion. Runs on stock ONNX Runtime. FlexAttention is replaced with an equivalent block-local + sink attention, and STFT/iSTFT are moved outside the model — I/O is `stft_in` / `masked_stft` `(1, 2, 1025, 801, 2)` (n_fft 2048, hop 441, 44.1kHz, 8-second chunks).
`wagiri/wagiri-roformer.onnx`	fp16, heavily optimized for WebGPU in wagiri. Not runnable on stock ONNX Runtime: the graph uses com.microsoft ops with app-specific semantics that require wagiri's patched onnxruntime-web kernels.
`wagiri/wagiri-roformer-fp32.onnx`	Same as above in fp32, for GPUs without `shader-f16`.

The conversion pipeline (with numerical-equivalence checks at every stage) lives in wagiri's conversion/.

MIT, same as the original code and weights.

Downloads last month: -; Downloads are not tracked for this model. How to track