File size: 10,812 Bytes
fac7535
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
#!/usr/bin/env python3
"""

inline_onnx.py

==============



把 `.model4` 里的 ONNX 图**内联成独立可跑的 .onnx**(不依赖 iZotope 运行时)。



背景

----

ONNX 图里大部分权重是 `external_data`,`location` 写成 `model4_<hex(张量名)>.bin`——

这些文件在磁盘上根本不存在,是 iZotope 的 Photon 在运行时从 safetensors 载荷里喂进去的。

所以抽出来的 .onnx 用 onnxruntime 打不开(`should be stored in model4_xxx.bin, but it is not regular file`)。



本脚本用容器自带的 `graphs[backend=onnx].bindings.tensors` 表把

`backend_name -> 容器张量名` 解析出来,把字节写回 initializer,输出标准 .onnx。



张量来源(三级回退)

--------------------

1. 容器里同名张量(规范名)

2. `bindings.tensors` 里的 `backend_name` 反查

3. `onnx.source.<sha>.<name>` 去掉前缀后的同名张量



用法

----

    python inline_onnx.py                 # 处理非 pymss 的 10 个模型

    python inline_onnx.py --all           # 连 pymss_models 里的 4 个也一起导

    python inline_onnx.py --only indoor

"""
from __future__ import annotations

import argparse
import json
import os
import re
import struct
import sys

import numpy as np
import onnx
import torch

import model_io
from onnx import numpy_helper

HERE = os.path.dirname(os.path.abspath(__file__))
SF_DIR = os.path.join(HERE, "..", "raw", "safetensors")
BLOBS_DIR = os.path.join(HERE, "..", "raw", "blobs")
CKPT_DIR = os.path.join(HERE, "..", "raw", "ckpt")
OUT_DIR = os.path.join(HERE, "..", "onnx")

# 已经塞进 pymss(stock BSRoformer 直接 strict 可加载)的 4 个
PYMSS_OK = {
    "voice2stem_v03_fp16c_t2", "voicesep2stem_fp16c_t2",
    "drumsep_1.3", "music4stemhq_fp16c_t2",
}


def read_header(model: str) -> dict:
    p = model_io.safetensors_path(model)
    with open(p, "rb") as f:
        n = struct.unpack("<Q", f.read(8))[0]
        hdr = json.loads(f.read(n))
    return json.loads(hdr["__metadata__"]["model4"])


def container_tensors(model: str) -> dict:
    sd = model_io.state_dict_for(model)
    out = {}
    for k, v in sd.items():
        if k.startswith("__blob__"):
            continue
        out[k] = v
        # 去掉 onnx.source.<sha>. / coreml.* 前缀后也登记一份
        m = re.match(r"^(?:onnx|coreml)\.source\.[0-9a-f]+\.(.*)$", k)
        if m:
            out.setdefault(m.group(1), v)
        m = re.match(r"^coreml\.fp16\.(.*)_to_fp16$", k)
        if m:
            out.setdefault(m.group(1), v)
    return out


def norm_name(k: str) -> str:
    n = k.replace("/", ".").replace("..", ".").strip(".")
    n = re.sub(r"^model\.", "", n)
    n = re.sub(r"^__onnx__\.", "", n)
    return n


def build_lookup(tensors: dict) -> dict:
    """归一化名 -> tensor,供名字只差分隔符/前缀时兜底。"""
    out = {}
    for k, v in tensors.items():
        out.setdefault(norm_name(k), v)
    return out


def binding_map(meta: dict) -> dict:
    """backend_name -> 容器张量名"""
    g = next((x for x in meta.get("graphs", []) if x.get("backend") == "onnx"), None)
    if not g:
        return {}
    out = {}
    for b in g.get("bindings") or []:
        for tname, v in (b.get("tensors") or {}).items():
            bn = v.get("backend_name")
            if bn and bn not in out:
                out[bn] = tname
    return out


def pick_onnx_blob(model: str):
    bd = model_io.blobs_dir(model)
    if not os.path.isdir(bd):
        return None
    cands = sorted((p for p in os.listdir(bd) if p.endswith(".onnx")),
                   key=lambda n: os.path.getsize(os.path.join(bd, n)), reverse=True)
    for n in cands:
        p = os.path.join(bd, n)
        try:
            m = onnx.load(p, load_external_data=False)
        except Exception:
            continue
        if len(m.graph.initializer) > 0:
            return p, m
    return None, None


def inline(model: str, tensors: dict, meta: dict):
    path, m = pick_onnx_blob(model)
    if m is None:
        return None, {"error": "没有可解析的 onnx blob"}

    bmap = binding_map(meta)
    look = build_lookup(tensors)
    # roformer 系(voice_restore 等)的 ONNX 权重存在规范名 blk.*/band_split.* 下,
    # 借用 to_pymss 的映射把 model.layers.* 还原回容器名
    try:
        sys.path.insert(0, HERE)
        from to_pymss import load_canon, map_bs_roformer
        ck = model_io.model_path_for(model)
        if ck and os.path.exists(ck):
            canon, al = load_canon(ck)
            if any(k.startswith("band_split") for k in canon):
                sd, _p, _u = map_bs_roformer(canon, al)
                for k, v in sd.items():
                    look.setdefault(norm_name("model." + k), v)
                    # 部分模型把注意力包在 .branch. 里
                    m2 = re.match(r"^(layers\.\d+\.\d+\.layers\.\d+\.\d+)\.(.+)$", k)
                    if m2:
                        look.setdefault(norm_name(f"model.{m2.group(1)}.branch.{m2.group(2)}"), v)
    except Exception as e:
        print("    (roformer 还原跳过:", e, ")")
    aliases = meta.get("tensor_aliases") or {}

    def deref(n):
        """沿 tensor_aliases 追到实际存储的张量名。"""
        seen = set()
        while isinstance(n, str) and n in aliases and n not in seen:
            seen.add(n)
            n = (aliases[n] or {}).get("tensor", n)
        return n
    need = [i for i in m.graph.initializer if any(e.key == "location" for e in i.external_data)]
    resolved = missing = 0
    for init in m.graph.initializer:
        loc = next((e.value for e in init.external_data if e.key == "location"), None)
        if loc is None:
            continue
        name = init.name
        t = tensors.get(name)
        if t is None and name in bmap:
            t = tensors.get(bmap[name])
        if t is None:
            try:
                dec = bytes.fromhex(loc.split("model4_")[-1].replace(".bin", "")).decode()
                t = tensors.get(dec) or tensors.get(bmap.get(dec, ""))
            except Exception:
                t = None
        if t is None:
            # Bandit:ONNX 里是 combined.0(LayerNorm) / combined.1(Linear),
            # 容器里是分开的 norm.* / fc.*
            m2 = re.match(r"^model\.(band_split\.norm_fc_modules\.\d+)\.combined\.([01])\.(\w+)$", name)
            if m2:
                sub = "norm" if m2.group(2) == "0" else "fc"
                t = look.get(norm_name(f"{m2.group(1)}.{sub}.{m2.group(3)}"))
        if t is None:
            t = look.get(norm_name(deref(name)))
        if t is None:
            t = look.get(norm_name(name))
        if t is None and name in bmap:
            # 关键一步:binding 的目标名往往还只是一层别名
            #   onnx.source.<sha>.model.band_split...  ->  band_split.0.norm.weight
            t = look.get(norm_name(deref(bmap[name])))
        if t is None and name in bmap:
            t = look.get(norm_name(bmap[name]))
        if t is None:
            missing += 1
            continue
        want = onnx.helper.tensor_dtype_to_np_dtype(init.data_type)
        arr = t.detach().cpu().numpy()
        if arr.dtype != want:
            arr = arr.astype(want, copy=False)
        if tuple(init.dims) and tuple(arr.shape) != tuple(init.dims):
            if arr.size != int(np.prod(init.dims)):
                missing += 1
                continue
            arr = arr.reshape(init.dims)
        init.CopyFrom(numpy_helper.from_array(arr, name=name))
        resolved += 1

    # 第二遍:首轮因解析顺序漏掉的再扫一次(按归一化名直接取)
    for init in m.graph.initializer:
        if not any(e.key == "location" for e in init.external_data):
            continue
        t2 = look.get(norm_name(init.name))
        if t2 is None:
            continue
        want2 = onnx.helper.tensor_dtype_to_np_dtype(init.data_type)
        arr2 = t2.detach().cpu().numpy()
        if arr2.dtype != want2:
            arr2 = arr2.astype(want2, copy=False)
        if tuple(init.dims) and tuple(arr2.shape) != tuple(init.dims):
            if arr2.size != int(np.prod(init.dims)):
                continue
            arr2 = arr2.reshape(init.dims)
        init.CopyFrom(numpy_helper.from_array(arr2, name=init.name))
        resolved += 1
        missing = max(0, missing - 1)

    os.makedirs(OUT_DIR, exist_ok=True)
    dst = os.path.join(OUT_DIR, model + ".onnx")   # 平铺,不套文件夹
    onnx.save(m, dst)
    info = {"model": model, "source_blob": os.path.basename(path),
            "nodes": len(m.graph.node), "initializers": len(m.graph.initializer),
            "external_refs": len(need), "resolved": resolved, "missing": missing,
            "bytes": os.path.getsize(dst)}
    # 校验
    try:
        onnx.checker.check_model(dst)
        info["checker"] = "OK"
    except Exception as e:
        info["checker"] = str(e)[:160]
    return dst, info


def main() -> int:
    ap = argparse.ArgumentParser()
    ap.add_argument("--all", action="store_true", help="包含已塞进 pymss 的 4 个")
    ap.add_argument("--only", default=None)
    args = ap.parse_args()
    os.makedirs(OUT_DIR, exist_ok=True)

    models = model_io.list_models()
    if not args.all:
        models = [n for n in models if n not in PYMSS_OK]
    if args.only:
        models = [n for n in models if args.only.lower() in n.lower()]

    report = []
    for model in models:
        try:
            meta = read_header(model)
            tensors = container_tensors(model)
            dst, info = inline(model, tensors, meta)
        except Exception as e:
            info = {"model": model, "error": f"{type(e).__name__}: {e}"}
            dst = None
        tag = "OK" if info.get("missing") == 0 and info.get("checker") == "OK" else "!!"
        print(f"[{tag}] {model:<32} " + (
            f"nodes={info.get('nodes')} inits={info.get('initializers')} "
            f"外部={info.get('external_refs')} 解析={info.get('resolved')} "
            f"缺={info.get('missing')} checker={info.get('checker')}"
            if "error" not in info else info["error"]))
        report.append(info)

    with open(os.path.join(OUT_DIR, "_inline_report.json"), "w", encoding="utf-8") as f:
        json.dump(report, f, indent=2, ensure_ascii=False)
    ok = sum(1 for r in report if r.get("missing") == 0 and r.get("checker") == "OK")
    print(f"\n完整内联 {ok}/{len(report)}")
    return 0


if __name__ == "__main__":
    sys.exit(main())