| """Initialize the decoder student by channel-slicing a frozen Vocos decoder.
|
|
|
| Training a decoder this small from scratch does not converge to intelligible
|
| audio; slicing the top-left corner of a pretrained Vocos vocoder's weights
|
| does. The Vocos backbone is 512-dim with 4 ConvNeXt blocks and pw 1536; this
|
| takes the first `--dim` channels, the first `--pw` hidden units and the first
|
| 3 blocks (dim=40 / pw=120 for the shipped model).
|
|
|
| The sliced checkpoint still needs the recovery + z-mix stages (train_decoder.py)
|
| before it is usable -- a fresh slice renders noise, and only the fine-tuning
|
| brings it back.
|
| """
|
| import torch
|
|
|
| import sys
|
| from pathlib import Path
|
|
|
| sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
|
| import torch.nn as nn
|
| from microtts.models import DecoderStudent
|
|
|
|
|
| def slice_linear(src, out_rows, in_cols):
|
| """Slice a [R, C] weight to [out_rows, in_cols] (top-left corner)."""
|
| return src[:out_rows, :in_cols].contiguous()
|
|
|
|
|
| def main():
|
| import argparse
|
| ap = argparse.ArgumentParser()
|
| ap.add_argument("--dim", type=int, default=40)
|
| ap.add_argument("--pw", type=int, default=120)
|
| ap.add_argument("--out", default="decoder_init.pt")
|
| args = ap.parse_args()
|
|
|
| from vocos import Vocos
|
| v = Vocos.from_pretrained("charactr/vocos-mel-24khz")
|
| sd = v.state_dict()
|
|
|
| student = DecoderStudent(dim=args.dim, pw_hidden=args.pw)
|
| ssd = student.state_dict()
|
|
|
|
|
| ssd["embed.weight"] = sd["backbone.embed.weight"][:args.dim].contiguous()
|
| ssd["embed.bias"] = sd["backbone.embed.bias"][:args.dim].contiguous()
|
|
|
| ssd["norm.weight"] = sd["backbone.norm.weight"][:args.dim].contiguous()
|
| ssd["norm.bias"] = sd["backbone.norm.bias"][:args.dim].contiguous()
|
|
|
| for i in range(3):
|
| t = f"backbone.convnext.{i}"
|
| s = f"blocks.{i}"
|
| ssd[f"{s}.gamma"] = sd[f"{t}.gamma"][:args.dim].contiguous()
|
| ssd[f"{s}.dw.weight"] = sd[f"{t}.dwconv.weight"][:args.dim].contiguous()
|
| ssd[f"{s}.dw.bias"] = sd[f"{t}.dwconv.bias"][:args.dim].contiguous()
|
| ssd[f"{s}.norm.weight"] = sd[f"{t}.norm.weight"][:args.dim].contiguous()
|
| ssd[f"{s}.norm.bias"] = sd[f"{t}.norm.bias"][:args.dim].contiguous()
|
| ssd[f"{s}.pw0.weight"] = slice_linear(sd[f"{t}.pwconv1.weight"], args.pw, args.dim)
|
| ssd[f"{s}.pw0.bias"] = sd[f"{t}.pwconv1.bias"][:args.pw].contiguous()
|
| ssd[f"{s}.pw1.weight"] = slice_linear(sd[f"{t}.pwconv2.weight"], args.dim, args.pw)
|
| ssd[f"{s}.pw1.bias"] = sd[f"{t}.pwconv2.bias"][:args.dim].contiguous()
|
|
|
| ssd["fnorm.weight"] = sd["backbone.norm.weight"][:args.dim].contiguous()
|
| ssd["fnorm.bias"] = sd["backbone.norm.bias"][:args.dim].contiguous()
|
|
|
| ssd["head.weight"] = slice_linear(sd["head.out.weight"], 1026, args.dim)
|
| ssd["head.bias"] = sd["head.out.bias"][:1026].contiguous()
|
|
|
| student.load_state_dict(ssd)
|
| torch.save(student.state_dict(), args.out)
|
| print(f"saved {args.out}")
|
| n = sum(p.numel() for p in student.parameters())
|
| print(f"student params: {n:,}")
|
|
|
|
|
| if __name__ == "__main__":
|
| main()
|
|
|
