#!/usr/bin/env python3 """ pymss_adapters.py ================= **不改 pymss 库代码**,在外部用适配类把 iZotope `.model4` 解出来的 ckpt 装进 pymss 模型。 适配思路 -------- pymss 的 `load_state_dict` 只认参数名,所以适配层要做两件事: 1. **补全张量**:iZotope 容器把一部分权重放在 `coreml.f16.*` / `coreml.source.*` 的扁平别名里 (例如 `model_mask_estimators_0_to_freqs_3_0_2_weight_to_fp16`),只从规范名取会缺一大半。 2. **对名**:把 iZotope 的命名规范化(去掉 `model.` `/` `.branch.`,`.`→`_`), 和 pymss 模型 `state_dict()` 的每个键做匹配;名字对不上的再用形状 + 出现顺序兜底。 三个模型各自的处理 ------------------ * `music4stemhq` : 架构与 bs_roformer 相同,只是参数多套了一层 `.branch.`(`BranchedRoformer`) * `musicnew4stem` : SCNet,命名只差一个 `model.` 前缀(`ScnetAdapter`) * `cinema4stem` : Bandit v2,`mask_estim` 是 ModuleDict(`BanditAdapter`) 用法 ---- python pymss_adapters.py # 三个模型全部转换 + strict 校验 """ from __future__ import annotations import argparse import json import os import re import sys import torch import model_io HERE = os.path.dirname(os.path.abspath(__file__)) SRC_DIR = os.path.join(HERE, "..", "raw") OUT_DIR = os.path.join(HERE, "..", "torch_models") PYMSS_CORE = r"E:\vs\pymss-core" # --------------------------------------------------------------------------- # # 张量收集:规范名 + 别名 # --------------------------------------------------------------------------- # def collect_tensors(path): """返回 {规范化名: tensor},兼容规范名 / coreml.source / coreml.fp16 三种来源。""" sd = model_io.load_any(path) raw = {} for k, v in sd.items(): if k.startswith(("__blob__", "__onnx__")): continue k = re.sub(r"^(coreml|onnx)\.source\.[0-9a-f]+\.", "", k) k = re.sub(r"^coreml\.fp16\.", "", k) if k.startswith(("buffer_", "/Constant", "Constant_")): continue if k.endswith("_to_fp16"): k = k[: -len("_to_fp16")] raw[k] = v out = {} for k, v in raw.items(): n = k.replace(".", "_").replace("/", "_").replace("_branch_", "_") n = re.sub(r"^model_", "", n) n = re.sub(r"_+", "_", n).strip("_") out[n] = v return out def norm_key(k: str) -> str: n = k.replace(".", "_").replace("/", "_").replace("_branch_", "_") n = re.sub(r"^model_", "", n) n = re.sub(r"_+", "_", n).strip("_") return n def match_state_dict(model, tensors): """把收集到的张量按规范化名对到 model 的 state_dict 上。""" ref = model.state_dict() by_name = {} unmatched = dict(tensors) for k in ref: n = norm_key(k) if n in unmatched: by_name[k] = unmatched.pop(n) # rotary_embed.freqs 是共享的确定性 buffer,广播到所有 (layer, axis) rot = next((v for k, v in by_name.items() if k.endswith("rotary_embed.freqs")), None) if rot is None: rot = next((v for nk, v in unmatched.items() if nk.endswith("rotary_embed_freqs")), None) if rot is not None: for k in ref: if k.endswith("rotary_embed.freqs") and k not in by_name: by_name[k] = rot # 名字没对上的,按形状 + 同名字段兜底 missing = [k for k in ref if k not in by_name] for k in list(missing): want = tuple(ref[k].shape) cand = [nk for nk, v in unmatched.items() if tuple(v.shape) == want and nk.split("_")[-1] == k.split(".")[-1]] if len(cand) == 1: by_name[k] = unmatched.pop(cand[0]) return by_name, [k for k in ref if k not in by_name], list(unmatched) # --------------------------------------------------------------------------- # # 适配类 # --------------------------------------------------------------------------- # class Adapter: """外部适配基类:持有 pymss 模型,负责建模型 / 装权重。""" name = "adapter" def __init__(self): self.model = None self.meta = {} def build(self): # pragma: no cover - 子类实现 raise NotImplementedError def load(self, ckpt): tensors = collect_tensors(ckpt) sd, missing, leftover = match_state_dict(self.model, tensors) # 直接改名装进去(不减 pymss 代码) st = self.model.state_dict() diff = [k for k, v in sd.items() if tuple(v.shape) != tuple(st[k].shape)] status = "OK" if not missing and not diff else f"缺 {len(missing)} / 形状不符 {len(diff)}" if not missing and not diff: self.model.load_state_dict(sd, strict=True) self.meta = {"model": self.name, "status": status, "matched": len(sd), "missing": missing[:8], "shape_mismatch": [f"{k} {tuple(st[k].shape)} vs {tuple(sd[k].shape)}" for k in diff[:8]], "leftover": len(leftover)} return self.meta class BranchedRoformer(Adapter): """music4stemhq:与 bs_roformer 同构,参数多一层 `.branch.`,由规范化名抹平。""" name = "music4stemhq_fp16c_t2" def build(self): from pymss_core.modules.bs_roformer.bs_roformer import BSRoformer from pymss_core.modules.bs_roformer.common import DEFAULT_FREQS_PER_BANDS self.model = BSRoformer(dim=256, depth=12, stereo=True, num_stems=4, time_transformer_depth=1, freq_transformer_depth=1, dim_head=64, heads=8, freqs_per_bands=DEFAULT_FREQS_PER_BANDS, stft_n_fft=2048, stft_hop_length=512, stft_win_length=2048, mask_estimator_depth=2, mlp_expansion_factor=4, use_shared_bias=False) return self def yaml(self): from to_pymss import yaml_bs_roformer from pymss_core.modules.bs_roformer.common import DEFAULT_FREQS_PER_BANDS return yaml_bs_roformer({"dim": 256, "depth": 12, "heads": 8, "dim_head": 64, "num_stems": 4, "mlp_expansion_factor": 4, "use_shared_bias": False, "freqs_per_bands": list(DEFAULT_FREQS_PER_BANDS)}, ["drums", "bass", "vocals", "other"]) class ScnetAdapter(Adapter): """musicnew4stem:SCNet,命名只差 `model.` 前缀。""" name = "musicnew4stem_fp16c_t2" def build(self): from pymss_core.modules.scnet.scnet import SCNet self.model = SCNet(dims=[4, 64, 128, 256], sources=["drums", "bass", "vocals", "other"], audio_channels=2, band_SR=[0.225, 0.372, 0.403], band_stride=[1, 4, 16], band_kernel=[3, 4, 16], conv_depths=[3, 2, 1], conv_kernel=3, num_dplayer=6, expand=1, compress=4, hop_size=1024, nfft=4096, win_size=4096, normalized=True) return self def yaml(self): return """audio: chunk_size: 352800 min_mean_abs: 0.0 num_channels: 2 sample_rate: 44100 inference: batch_size: 1 dim_t: 256 normalize: false num_overlap: 2 model: audio_channels: 2 band_SR: - 0.225 - 0.372 - 0.403 band_kernel: - 3 - 4 - 16 band_stride: - 1 - 4 - 16 compress: 4 conv_depths: - 3 - 2 - 1 conv_kernel: 3 dims: - 4 - 64 - 128 - 256 expand: 1 hop_size: 1024 nfft: 4096 normalized: true num_dplayer: 6 sources: - drums - bass - other - vocals win_size: 4096 model_type: scnet training: instruments: - drums - bass - vocals - other """ class BanditAdapter(Adapter): """cinema4stem:Bandit v2,mask_estim 是 ModuleDict(music/sfx/speech)。""" name = "cinema4stem_fp16c_t2" def build(self): from pymss_core.modules.bandit_v2.bandit import Bandit self.model = Bandit(in_channels=2, stems=["music", "sfx", "speech"], fs=44100, band_type="musical", n_bands=64, n_sqm_modules=12, emb_dim=128, rnn_dim=256, bidirectional=True, rnn_type="GRU", mlp_dim=512, hidden_activation="Tanh", complex_mask=True, use_freq_weights=True, n_fft=2048, win_length=2048, hop_length=512) return self def yaml(self): return """audio: chunk_size: 352800 min_mean_abs: 0.0 num_channels: 2 sample_rate: 44100 inference: batch_size: 1 dim_t: 256 num_overlap: 2 model: band_type: musical bidirectional: true complex_mask: true emb_dim: 128 fs: 44100 hidden_activation: Tanh in_channels: 2 mlp_dim: 512 n_bands: 64 n_fft: 2048 n_sqm_modules: 12 hop_length: 512 win_length: 2048 rnn_dim: 256 rnn_type: GRU stems: - music - sfx - speech use_freq_weights: true model_type: bandit training: instruments: - dialogue - music - sfx """ ADAPTERS = [BranchedRoformer, ScnetAdapter, BanditAdapter] def main() -> int: ap = argparse.ArgumentParser() ap.add_argument("--src", default=SRC_DIR) ap.add_argument("--out", default=OUT_DIR) args = ap.parse_args() sys.path.insert(0, PYMSS_CORE) sys.path.insert(0, HERE) os.makedirs(args.out, exist_ok=True) report = [] for cls in ADAPTERS: a = cls().build() ckpt = model_io.model_path_for(a.name) or os.path.join(args.src, a.name, a.name + ".ckpt") if not os.path.exists(ckpt): print(f"[{a.name}] 缺少 ckpt") continue meta = a.load(ckpt) d = os.path.join(args.out, a.name) os.makedirs(d, exist_ok=True) with open(os.path.join(d, "config.yaml"), "w", encoding="utf-8") as f: f.write(a.yaml()) if meta["status"] == "OK": sd = a.model.state_dict() bad = [k for k, v in sd.items() if "weight" not in k and "bias" not in k and "gamma" not in k and "freqs" not in k] torch.save({"state_dict": {k: v.clone() for k, v in sd.items()}}, os.path.join(d, a.name + ".ckpt")) print(f"[{a.name}] {meta['status']} 匹配 {meta['matched']} 多余 {meta['leftover']}") if meta["missing"]: print(" 缺:", meta["missing"]) if meta["shape_mismatch"]: print(" 形状不符:", meta["shape_mismatch"]) report.append(meta) with open(os.path.join(args.out, "_adapters_report.json"), "w", encoding="utf-8") as f: json.dump(report, f, indent=2, ensure_ascii=False) return 0 if __name__ == "__main__": sys.exit(main())