anira-project/rave-djembe

RAVE djembe for anira

A RAVE model of djembe (trained by Michael Witte) with all streaming state as explicit tensors: encoder, decoder and forward for ONNX Runtime, LiteRT and ExecuTorch, 1024 samples per call at 44.1 kHz, plus the original streaming TorchScript export.

Files

File	Engine	Size	SHA256
onnx/rave_decoder.onnx	ONNX Runtime	45,194,034	74028e7746b6922c6d56943389b9d79789a9bc7fd9aaaa9823b532e68b2400fb
onnx/rave_encoder.onnx	ONNX Runtime	10,124,410	93cb3ae7dd6b7c1ff024e87bf019ed7ca02e31de00ab7693bad9c2ee6518e6e6
onnx/rave_forward.onnx	ONNX Runtime	55,318,855	622b79a4aa1fe43770db4653fb016b7976f100854e9c0bc4039a41ed49c975c3
litert/rave_decoder.tflite	LiteRT, TensorFlow Lite	45,275,288	3e53d2fe90fbd9a2f5f53c106ad81a71f232bdb00c3b9408e4d9d5d60a3ca77b
litert/rave_encoder.tflite	LiteRT, TensorFlow Lite	10,153,184	4c734db596589758d161bc8f6fb542dce62fab89914cfdd1cd68442e87a3488f
litert/rave_forward.tflite	LiteRT, TensorFlow Lite	55,414,580	7c6de2b0523e1006b70398da157fc9cd81f75db2a208dbe4f0b4b076aa8e518d
libtorch/Djembe_streaming.ts	LibTorch	96,819,779	33a8a64a8e90043f147e00effcdfa06f21dfb8e935fc3f7c568fcc8d30a06483
libtorch/rave_decoder.pt	LibTorch	45,559,732	c87cc32fbd3488405984f9617d36e80b6ee2209f1fa617d480ef0081540aa7ac
libtorch/rave_encoder.pt	LibTorch	10,195,660	75e7d521da6c07235bb2ffcf5e58d8be1aabdedffb0b6dccac80a4dc2c50dfbb
libtorch/rave_forward.pt	LibTorch	55,752,892	35c6d69937519cfe1475367e51a9ce81cfb8f3cda1ae6c0080c1d9119873fee9
executorch/rave_decoder.pte	ExecuTorch 1.3.1, CPU	84,714,240	b5a9dadf852fc3e25579db286a66b602b90ade799595ae81275ec167eb635505
executorch/rave_decoder_coreml.pte	ExecuTorch 1.3.1, Core ML (Apple)	45,652,485	fcf21307a1c0b509d7e11ea594589eb0292269e4bc9403dd1429b7f20f366bbf
executorch/rave_decoder_vulkan.pte	ExecuTorch 1.3.1, Vulkan (GPUs)	45,123,648	84ee3c7d2eb4d5e94c57f87c5515e4c031b960ba0061774b8e9f63bcc5a63350
executorch/rave_encoder.pte	ExecuTorch 1.3.1, CPU	11,994,112	c25870b0f3783007a6ecf666fd215829d99e3369a7a7c2bf273cbdd2c02ae6be
executorch/rave_encoder_coreml.pte	ExecuTorch 1.3.1, Core ML (Apple)	10,218,569	145d900e73739727d53d9e6349edf052240f5419e5daadf43469431396bdbc6d
executorch/rave_encoder_vulkan.pte	ExecuTorch 1.3.1, Vulkan (GPUs)	10,122,688	4a828d495144fc6ea31943a946eb3971c3002e5787fb2744ab642a6aa582940f
executorch/rave_forward.pte	ExecuTorch 1.3.1, CPU	96,626,816	ed328d7e422334b95222e3998592b09e97eef1814ee3d0edcfa82d8b639808e7
executorch/rave_forward_coreml.pte	ExecuTorch 1.3.1, Core ML (Apple)	55,862,232	68583a4592b5e8b2a3bfd203ae7b2992b42de53f29b9e9d1b3afa420cec9217d
executorch/rave_forward_vulkan.pte	ExecuTorch 1.3.1, Vulkan (GPUs)	55,241,408	dd9230f71184e28ecb6f44f27b98558ede083056421342f96e70e9163cf34ff3
coreml/rave_decoder.mlpackage	Core ML (native)	45,188,865	per file in the catalog
coreml/rave_encoder.mlpackage	Core ML (native)	10,134,979	per file in the catalog
coreml/rave_forward.mlpackage	Core ML (native)	55,317,954	per file in the catalog

Source

anira-project/example-models @ d1bc5cf, third-party/ircam-acids/RAVE/RaveDjembe/; the export pipeline and its parity tests are there. The model was developed and trained by Michael Witte with acids-ircam/RAVE.

Formats and checks

Every model here runs on ONNX Runtime, LiteRT and TensorFlow Lite, LibTorch, ExecuTorch (XNNPACK or portable kernels on the CPU, Core ML on Apple, Vulkan on GPUs) and as a native Core ML ML program, with the ONNX models' tensor names. anira foundation models streams every format of every description over twelve consecutive calls with the state fed back and checks that they agree (tools/check_streaming.py). Deviations below are from ONNX Runtime, per model, in the order of the files.

Format	Files	Max deviation
LibTorch (traced, explicit state)	libtorch/*.pt	5.6e-06, 5.1e-08, 4.7e-06
ExecuTorch, Core ML (Apple, float32)	executorch/*_coreml.pte	5.5e-06, 1.3e-07, 8.3e-06
ExecuTorch, Vulkan (GPUs)	executorch/*_vulkan.pte	exported; not run here (no Vulkan runtime)
Core ML ML program (anira's native engine, coming with v3; macOS 13, iOS 16)	coreml/*.mlpackage	5.5e-06, 1.3e-07, 8.3e-06

uv run --no-project export/export_litert.py <model.onnx> <out.tflite> [<description>] [--guitar-lstm <pt>]
uv run --no-project export/export_executorch.py <model.onnx> <out.pte> [<description>] [--guitar-lstm <pt>] [--coreml | --vulkan]
uv run --no-project export/export_coreml.py <model.onnx> <out.mlpackage> [<description>] [--guitar-lstm <pt>]
uv run --no-project export/export_libtorch.py <model.onnx> <out.pt> [<description>] [--guitar-lstm <pt>]

<description> is the model's description in anira foundation models, for the input shapes where the ONNX model's time axis is dynamic. The PyTorch module comes from the ONNX model (export/onnx_to_torch.py: onnxsim to static shapes, onnx2torch with traceable converters); each export is checked against ONNX Runtime over eight consecutive calls of random input with the state fed back, relative to the output's peak (export/check.py).

This repository holds the weights, the export scripts and the licence. The anira model descriptions live with the code in anira foundation models (models/rave-djembe), versioned with git. Pin a revision and check each file's SHA256 when you download:

https://huggingface.co/anira-project/rave-djembe/resolve/<revision>/onnx/rave_decoder.onnx

Licence

CC BY-NC 4.0, non-commercial use only: LICENSE, copied from the source repository.