Repository navigation
Expand file tree
/
Copy pathrun_sample_exp.py
More file actions
373 lines (310 loc) · 12.4 KB
/
Copy pathrun_sample_exp.py
File metadata and controls
373 lines (310 loc) · 12.4 KB
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
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
单样本音乐编辑实验脚本
--------------------------
基于 AudioLDM2 与 Polyphonia 编辑框架,
针对单个 sample/ 目录执行免训练多轨音乐编辑。
# 基础运行(编辑 sample/ 下所有任务)
python run_sample_exp.py --sample_dir ./sample --output_dir ./sample/output --visualize
# 仅执行 vocals2violin 任务
python run_sample_exp.py --task_key vocals2violin --device cuda --seed 928
"""
from __future__ import annotations
import argparse
import copy
import gc
import json
import logging
import os
import random
import sys
from pathlib import Path
from typing import Any, Dict, List, Optional
import numpy as np
import torch
from audio_utils import generate_soft_masks
from edit import Edit
from my_audioldm2 import load_audioldm2
from utils import visualize_and_save_spectrogram
logger = logging.getLogger("run_sample_exp")
def setup_logging() -> None:
logger.handlers.clear()
logger.setLevel(logging.INFO)
console_handler = logging.StreamHandler(sys.stdout)
console_handler.setFormatter(logging.Formatter(
"[%(asctime)s][%(levelname)s] %(message)s", datefmt="%Y-%m-%d %H:%M:%S"
))
logger.addHandler(console_handler)
logging.getLogger("transformers").setLevel(logging.ERROR)
logging.getLogger("diffusers").setLevel(logging.ERROR)
logging.getLogger("audioldm2").setLevel(logging.WARNING)
BEST_EDIT_PARAMS: Dict[str, Any] = {
"steps": 100,
"use_attention_mask": False,
"tgt_block": 2,
"swap_break_point": 0,
"guidance_scale": 3.5,
"t_start": 1000,
"mask_bias_strength": 2.5,
"self_layers": {
"down": [i for i in range(1, 19)],
"up": [],
"mid": [],
},
"cross_layers": {
"down": [i for i in range(1, 19)],
"up": [],
"mid": [],
},
}
DEFAULT_STFT_PARAMS = {
"sampling_rate": 16000,
"n_fft": 1024,
"hop_length": 160,
"win_length": 1024,
"n_mels": 64,
"mel_fmin": 0.0,
"mel_fmax": 8000.0,
}
def select_device(device_arg: str) -> str:
if device_arg == "cuda":
return "cuda"
if device_arg == "cpu":
return "cpu"
return "cuda" if torch.cuda.is_available() else "cpu"
def select_precision(precision_arg: str, device: str) -> torch.dtype:
if precision_arg == "fp16":
return torch.float16
if precision_arg == "fp32":
return torch.float32
return torch.float16 if device == "cuda" else torch.float32
def parse_prompt_file(prompt_json: Path, task_key: Optional[str] = None) -> Dict[str, Any]:
try:
with open(prompt_json, "r", encoding="utf-8") as f:
data = json.load(f)
except Exception as exc:
logger.warning(f"读取 prompt 失败: {prompt_json}, err={exc}")
return {}
if not isinstance(data, dict) or not data:
return {}
tasks: Dict[str, Dict[str, Any]] = {}
key_mapping = {
"Vocals Softmask Prompt": "vocals",
"Bass Softmask Prompt": "bass",
"Drums Softmask Prompt": "drums",
"Others Softmask Prompt": "other",
"Other Softmask Prompt": "other",
}
for task_name, block in data.items():
if not isinstance(block, dict):
continue
baseline = block.get("Baseline Prompt", "") or ""
target_prompts: Dict[str, str] = {}
for src_key, alias in key_mapping.items():
val = block.get(src_key, "")
if isinstance(val, str) and val.strip():
target_prompts[alias] = val.strip()
tasks[task_name] = {
"baseline_prompt": baseline.strip(),
"target_prompt": target_prompts,
"raw": block,
}
if task_key:
tasks = {task_key: tasks[task_key]} if task_key in tasks else {}
return tasks
def infer_edit_stem(task_name: str) -> List[str]:
"""
根据任务名称推断应该被编辑的乐器分轨。
例如:"vocals2violin" -> ["vocals"], "drums2percussion" -> ["drums"]
"""
name = (task_name or "").lower()
stems: List[str] = []
if "vocals2" in name:
stems.append("vocals")
if "drums2" in name:
stems.append("drums")
if "bass2" in name:
stems.append("bass")
if not stems:
stems.append("other")
return stems
def save_audio_safe(path: Path, audio_np: np.ndarray, sr: int = 16000) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
try:
import soundfile as sf
sf.write(str(path), audio_np, sr)
except Exception:
try:
import scipy.io.wavfile as wavfile
wavfile.write(str(path), sr, (audio_np * 32767).astype(np.int16))
except Exception as e:
logger.error(f"[保存失败] {path}: {e}")
def reconfigure_editor(ed: Edit, params: Dict[str, Any], preserve_cache: bool = False) -> None:
setattr(ed, "_preserve_attn_cache", preserve_cache)
ed.unsetting()
ed.edit_params_default.update(params)
ed.t_start = params.get("t_start", ed.t_start)
ed.swap_break_point = params.get("swap_break_point", ed.swap_break_point)
if "mask_bias_strength" in params:
ed.mask_bias_strength = params["mask_bias_strength"]
ed.setting()
setattr(ed, "_preserve_attn_cache", False)
def prepare_editor(device: str, precision_t: torch.dtype, init_params: Dict[str, Any]) -> Edit:
logger.info("正在加载 AudioLDM2 模型...")
vae, tokenizer, tokenizer_2, text_encoder, text_encoder_2, vocoder, projection_model, language_model, unet, scheduler = load_audioldm2(
precision_t=precision_t, device=device
)
ed = Edit(
vae, tokenizer, tokenizer_2, text_encoder, text_encoder_2,
vocoder, projection_model, language_model, unet, scheduler,
device, edit_params=copy.deepcopy(init_params),
)
ed.edit_params_default.update({"t_start": init_params["t_start"]})
ed.t_start = init_params["t_start"]
ed.mask_bias_strength = init_params["mask_bias_strength"]
ed.swap_break_point = init_params["swap_break_point"]
return ed
def build_prompt_for_generation(prompt_cfg: Dict[str, Any]) -> Any:
baseline = prompt_cfg.get("baseline_prompt", "") or ""
target = prompt_cfg.get("target_prompt", {}) or {}
if baseline.strip() and isinstance(target, dict) and target:
prompt = {"__baseline__": baseline.strip()}
prompt.update(target)
return prompt
if baseline.strip():
return baseline.strip()
if isinstance(target, dict) and target:
prompt = {"__baseline__": ""}
prompt.update(target)
return prompt
return ""
def main() -> None:
parser = argparse.ArgumentParser(description="单样本音乐编辑实验脚本")
parser.add_argument("--sample_dir", type=str, default="./sample",
help="输入样本目录,需包含 mixture.wav、prompt_multi.json 及各分轨 wav")
parser.add_argument("--output_dir", type=str, default="./sample/output",
help="输出结果目录")
parser.add_argument("--device", type=str, default="auto",
help="设备选择: cuda/cpu/auto")
parser.add_argument("--precision", type=str, default="auto",
help="精度选择: fp16/fp32/auto")
parser.add_argument("--seed", type=int, default=928,
help="随机种子")
parser.add_argument("--visualize", action="store_true",
help="是否保存频谱图可视化")
parser.add_argument("--mask_type", type=str, default="energy",
help="软掩码类型")
parser.add_argument("--task_key", type=str, default=None,
help="可选:指定 prompt_multi.json 中的任务键,仅执行该任务")
args = parser.parse_args()
sample_dir = Path(args.sample_dir).expanduser().resolve()
output_dir = Path(args.output_dir).expanduser().resolve()
mixture_wav = sample_dir / "mixture.wav"
prompt_json = sample_dir / "prompt_multi.json"
if not mixture_wav.exists():
logger.error(f"mixture.wav 不存在: {mixture_wav}")
return
if not prompt_json.exists():
logger.error(f"prompt_multi.json 不存在: {prompt_json}")
return
setup_logging()
torch.manual_seed(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(args.seed)
device = select_device(args.device)
precision_t = select_precision(args.precision, device)
logger.info("========== 单样本音乐编辑启动 ==========")
logger.info(f"样本目录: {sample_dir}")
logger.info(f"输出目录: {output_dir}")
logger.info(f"设备/精度: {device} / {precision_t}")
logger.info(f"随机种子: {args.seed}")
init_params = copy.deepcopy(BEST_EDIT_PARAMS)
try:
ed = prepare_editor(device, precision_t, init_params)
except Exception as exc:
logger.critical(f"模型加载失败: {exc}", exc_info=True)
return
# 解析 prompt_multi.json
prompt_tasks = parse_prompt_file(prompt_json, args.task_key)
if not prompt_tasks:
logger.error("未解析到有效的编辑任务,退出。")
return
logger.info(f"发现 {len(prompt_tasks)} 个编辑任务: {list(prompt_tasks.keys())}")
# 生成软掩码
try:
masks = generate_soft_masks(
str(sample_dir), None, DEFAULT_STFT_PARAMS, 0.0, args.mask_type
)
except Exception as exc:
logger.error(f"软掩码生成失败: {exc}", exc_info=True)
return
if not masks:
logger.error("软掩码为空,退出。")
return
logger.info(f"软掩码已生成,包含分轨: {list(masks.keys())}")
# 反演: 对 mixture.wav 执行一次 Inversion 获取初始 Latent
try:
reconfigure_editor(ed, init_params, preserve_cache=False)
ed.set_soft_masks(masks)
ed.enable_softmasking(True)
ed.trigger_cross_related = True
ed.trigger_self_related = True
ed.trigger_get_map, ed.trigger_swap_map = True, False
mel, _, _, audio_len = ed.load_wav(str(mixture_wav), None, ed.device)
inv_latents = ed.invert_pipeline(mel, audio_len, prompt="")
inv_latent = inv_latents[-1]
ed.trigger_get_map = False
logger.info(f"反演完成,音频长度: {audio_len:.2f}s")
except Exception as exc:
logger.error(f"反演失败: {exc}", exc_info=True)
return
# 配置编辑器用于生成(保留反演阶段缓存的注意力图)
reconfigure_editor(ed, init_params, preserve_cache=True)
ed.set_soft_masks(masks)
ed.enable_softmasking(True)
ed.trigger_cross_related = True
ed.trigger_self_related = True
# 遍历每个任务执行 Reverse 生成
success_count = 0
fail_count = 0
for task_name, prompt_cfg in prompt_tasks.items():
logger.info(f"--> 执行任务: {task_name}")
out_wav = output_dir / f"{task_name}.wav"
if out_wav.exists() and out_wav.stat().st_size > 1024:
logger.info(f"[跳过] 已存在: {out_wav}")
success_count += 1
continue
try:
edit_stems = infer_edit_stem(task_name)
ed.set_edit_stem(edit_stems)
ed.trigger_get_map = False
ed.trigger_swap_map = True
prompt_to_use = build_prompt_for_generation(prompt_cfg)
last_mel = ed.reverse_pipeline(
prompt=prompt_to_use, audio_length=audio_len, init_latent=inv_latent
)
audio_np = ed.mel_spectrogram_to_audio(last_mel).audios[0]
save_audio_safe(out_wav, audio_np)
logger.info(f"[成功] 已保存音频: {out_wav}")
if args.visualize:
out_png = output_dir / f"{task_name}.png"
visualize_and_save_spectrogram(
out_wav, out_png, DEFAULT_STFT_PARAMS,
title=f"{task_name}"
)
logger.info(f"[成功] 已保存频谱图: {out_png}")
success_count += 1
except Exception as exc:
logger.error(f"[失败] 任务 {task_name}: {exc}", exc_info=True)
fail_count += 1
finally:
if torch.cuda.is_available():
torch.cuda.empty_cache()
gc.collect()
logger.info("========== 单样本音乐编辑完成 ==========")
logger.info(f"成功: {success_count} | 失败: {fail_count}")
if __name__ == "__main__":
main()