"""Converte `best.pt` (PyTorch) e os ONNX do vocoder em pesos MLX (safetensors). Uso: python code/convert.py pivo-ptbr.pt onnx mlx onnx/metadata.json """ import json, sys from pathlib import Path import numpy as np import onnx import torch from onnx import numpy_helper as nh from safetensors.numpy import save_file def acoustic(ckpt): sd = {k: v.numpy().astype(np.float32) for k, v in ckpt["model"].items()} out = {} for k, v in sd.items(): if v.ndim == 3: # Conv1d (out,in/groups,k) -> MLX (out,k,in/groups) v = v.transpose(0, 2, 1) if ".dw." in k or ".pw." in k or ".norms." in k: pass out[k] = np.ascontiguousarray(v) return out def vocoder(directory): d = Path(directory) c = {i.name: nh.to_array(i) for i in onnx.load(d / "conditioning.onnx").graph.initializer} s = {i.name: nh.to_array(i) for i in onnx.load(d / "vocoder_step.onnx").graph.initializer} m = lambda n: s[f"onnx::MatMul_{n}"] w = {"cond.pitch_embedding": c["pitch_embedding.weight"], "cond.dense1": c["onnx::MatMul_20"], "cond.conv": c["conv.weight"].transpose(0, 2, 1), "cond.dense2": c["onnx::MatMul_21"], "cond_gain.w": s["cond_gain.weight"], "cond_gain.b": s["cond_gain.bias"], "pitch_gain.w": s["pitch_gain.weight"], "pitch_gain.b": s["pitch_gain.bias"], "frame_conv": m(274), "frame_glu": m(275), "gru1.input": m(282), "gru1.hidden": m(283), "glu1": m(286), "gru2.input": m(293), "gru2.hidden": m(294), "glu2": m(297), "gru3.input": m(304), "gru3.hidden": m(305), "glu3": m(308), "skip_dense": m(315), "skip_glu": m(318), "signal_out": m(319)} return {k: np.ascontiguousarray(v.astype(np.float32)) for k, v in w.items()} if __name__ == "__main__": ckpt_path, vdir, out = sys.argv[1:4] out = Path(out); out.mkdir(parents=True, exist_ok=True) ck = torch.load(ckpt_path, map_location="cpu", weights_only=True) save_file(acoustic(ck), str(out / "acoustic.safetensors")) save_file(vocoder(vdir), str(out / "vocoder.safetensors")) meta = json.loads(Path(sys.argv[4] if len(sys.argv) > 4 else Path(vdir) / "metadata.json").read_text()) meta["format"] = "pivo-ptbr-mlx-v1"; meta["config"] = {**ck["config"]} (out / "config.json").write_text(json.dumps({"config": ck["config"], "sample_rate": 16000, "step": ck["step"], "voices": meta["voices"], "symbols": meta["symbols"], "preemphasis": 0.85, "max_utterance_frames": 3000, "format": "pivo-ptbr-mlx-v1"}, ensure_ascii=False, indent=1)) print("ok", out)