aaatmy/game-large-onnx

GAME Large — official v1.0.3 ONNX mirror

Community mirror maintained by Hikari Tsai (aaatmy), not an official openvpi repository. The five FP32 ONNX graphs and config.json are copied without modification from openvpi/GAME v1.0.3. These are the official ONNX exports of the GAME 1.0 Large pretrained weights (~100M parameters, embedding dim 256, 8 heads, 8+16+8 layers), not a newly trained model or a quantized conversion. The Small variant is mirrored at https://huggingface.co/aaatmy/game-small-onnx.

• Hugging Face: https://huggingface.co/aaatmy/game-large-onnx
• ONNX files: 393,801,733 bytes total (393.8 MB), opset 17, trace exporter.
• Source ZIP: GAME-1.0.3-large-onnx.zip, 361,619,205 bytes.
• Source ZIP SHA-256: 8a5480539fe7d995800dc0efe149b83a1cd4f4e4a36aa2a1f7665f1765dcac08.

Files and usage

File	Purpose
encoder.onnx	Encode the audio waveform into model features (waveform [B, L], duration [B] → x_seg, x_est, maskT)
segmenter.onnx	Iteratively predict note boundaries (D3PM loop over t)
estimator.onnx	Estimate note pitches (scores, MIDI semitones) and presence flags
dur2bd.onnx	Convert known durations to boundaries
bd2dur.onnx	Convert boundaries to durations
config.json	Model configuration, sample rate and language IDs
manifest.json, SHA256SUMS	Upstream provenance, file sizes and SHA-256 digests

Use the graphs together as described in the versioned upstream ONNX documentation. This is not a Hugging Face Transformers or Diffusers pipeline. It expects a host implementation of preprocessing, iterative segmentation and result decoding. The configuration uses 44,100 Hz audio and a 0.01-second time step.

Notes for ONNX Runtime hosts: encoder.onnx contains an STFT node and Einsum nodes, and segmenter.onnx contains a RandomUniformLike node for D3PM sampling. All graphs load and run on the CPU execution provider (tested with onnxruntime 1.31). Execution providers that do not implement these ops fall back to CPU for those nodes. Segmenter cost grows faster than linearly with input length, so long recordings should be processed in chunks.

To download the complete bundle:

hf download aaatmy/game-large-onnx --local-dir ./game-large-onnx

For reproducible applications, add --revision <commit-sha> to pin the model snapshot. To check the original assets after download, run shasum -a 256 -c SHA256SUMS inside the downloaded directory (or sha256sum -c SHA256SUMS on Linux).

Attribution and license

Attribution as required by CC BY-NC-SA 4.0 (title, author, source, license, changes):

• Title: GAME 1.0 Large (singing-voice MIDI extraction model, ONNX export v1.0.3)
• Author / licensor: Team OpenVPI (the openvpi/GAME project; models published by yqzhishen; data labels by 挚彬Club, YQ之神, 小狼 and 幽寂)
• Source: https://github.com/openvpi/GAME/releases/tag/v1.0.3 (training sources and acknowledgements in the v1.0.0 release)
• License: CC BY-NC-SA 4.0
• Changes: none to the model files; this mirror adds manifest.json, SHA256SUMS and this card.

The model files are licensed under CC BY-NC-SA 4.0, as specified in that release. The full legal text is included in LICENSE: attribution, noncommercial use and ShareAlike conditions apply. This mirror adds only documentation and checksums; the model files and configuration are unchanged. Any future conversion or quantization must retain the applicable model terms and identify its changes.

The upstream code repository uses MIT, but its MIT license does not replace the model weights' CC BY-NC-SA 4.0 license. The models are provided without warranties under their license; no upstream endorsement of this mirror is implied.

繁體中文

此處託管官方 GAME Large v1.0.3 的原始 FP32 ONNX，共五個模型及設定檔， 總模型大小約 393.8 MB，未進行量化或修改。由 Hikari Tsai 建立社群副本， 並非 openvpi 官方 Hugging Face 帳戶。Small 版本見 aaatmy/game-small-onnx。

模型作者為 Team OpenVPI（openvpi/GAME，由 yqzhishen 發布），採 CC BY-NC-SA 4.0（姓名標示、非商業性、相同方式分享）。 來源、檔案大小與校驗碼列於 manifest.json。