File size: 1,246 Bytes
7688e10 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 | {
"architecture": "HTDemucs (demucs 4.1.0, MIT)",
"sources": [
"drums",
"bass",
"other",
"vocals"
],
"samplerate": 44100,
"audio_channels": 2,
"model_cfg": {
"channels": 44,
"t_layers": 5,
"t_heads": 8,
"bottom_channels": 0,
"segment": 6.0
},
"model_name": "medium",
"nparams": 22612960,
"checkpoint": "best",
"step": 120450,
"valid_sdr": {
"drums": 11.869,
"bass": 8.689,
"other": 2.352,
"vocals": 10.811,
"mean": 8.43,
"n_songs": 44
},
"valid_sdr_mix": {
"drums": 7.507,
"bass": 8.648,
"other": 0.749,
"vocals": 5.304,
"mean": 5.552,
"n_songs": 41
},
"provenance": "own weights: random init, trained on the user's Suno stems (sum-of-stems mixtures); stage 2: fine-tuned from our stage-1 weights (best.pt, step 85702, sha256 ad21ed8115643425\u2026) with Suno's delivered mixes as part of the inputs and a multi-resolution STFT loss",
"load": "separation_train.build_model(model_cfg, sources).load_state_dict(torch.load('weights.pth'))",
"input_normalisation": "x = (mix - mean(mono)) / std(mono); output * std + mean (as demucs apply)",
"onnx": {
"file": "model_fp16.onnx",
"input": "mix",
"shape": [
1,
4,
2,
264600
],
"max_abs_diff": 0.0002949833869934082
}
} |