File size: 13,033 Bytes
bc7b7ea fcf8a1d bc7b7ea fcf8a1d bc7b7ea | 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 | """Evaluate every Ming TTS tier on the fixed WebFM eval set (eval/sentences.json).
Runs the teacher and the distilled students over the same voice x sentence
matrix, writes one wav per cell and one JSONL row per cell, then leaves the
scoring to `eval/score.py`.
The teacher is generated in the same run and on the same sentences: a candidate
is only meaningful as a delta against it.
Expected Kaggle inputs:
/kaggle/input/webfm-eval/eval/sentences.json (this repo's eval/)
/kaggle/input/webfm-eval/config/ming_voices.json
/kaggle/input/webfm-eval/refs/wav/v01.wav .. v08.wav
/kaggle/working/Ming-3.0-tts + Ming-3.0-tts-0.5b (as in ming_multistep.py)
Env:
VOICES=v01,v02,... subset of voices, for splitting across two GPUs
MODELS=teacher,3step,4step
LIMIT=40 sentences to run
SEED=20261021
Resumable: a cell whose wav exists is not regenerated; the JSONL is appended
and flushed per row, so a killed session keeps what it finished.
"""
from __future__ import annotations
import json
import os
import re
import sys
import time
import traceback
from pathlib import Path
import numpy as np
import torch
import torchaudio
ROOT = Path("/kaggle/working")
REPO = ROOT / "Ming-3.0-tts"
MODEL_DIR = ROOT / "Ming-3.0-tts-0.5b"
DATA = Path(os.environ.get("EVAL_DATA", "/kaggle/input/webfm-eval"))
OUT = ROOT / "ming_eval"
WAV = OUT / "wav"
ROWS = OUT / "results.jsonl"
LOG = OUT / "eval.log"
SEED = int(os.environ.get("SEED", "20261021"))
LIMIT = int(os.environ.get("LIMIT", "40"))
VOICE_FILTER = [v for v in os.environ.get("VOICES", "").split(",") if v]
MODELS = [m for m in os.environ.get("MODELS", "teacher,3step,4step").split(",") if m]
TEACHER_STEPS = 10
TEACHER_CFG = 2.0
STUDENT_REPOS = {
"1step": ("idefinity/ming-3.0-tts-1step-distill-int4", 1),
"2step": ("idefinity/ming-3.0-tts-2step-distill-int4", 2),
"3step": ("idefinity/ming-3.0-tts-3step-distill-int4", 3),
"4step": ("idefinity/ming-3.0-tts-4step-distill-int4", 4),
}
for p in (OUT, WAV):
p.mkdir(parents=True, exist_ok=True)
def log(msg: str) -> None:
line = f"[{time.strftime('%H:%M:%S')}] {msg}"
print(line, flush=True)
with open(LOG, "a") as f:
f.write(line + "\n")
def patch_fp32() -> None:
"""T4 has no bf16; the repo defaults to bf16 and NaNs out under fp16."""
for name in ("ming_tts.py", "modeling_bailingmm.py"):
path = REPO / name
s = path.read_text()
s = s.replace("torch_dtype=torch.bfloat16", "torch_dtype=torch.float32")
s = s.replace(".eval().to(torch.bfloat16).to(device)", ".eval().to(torch.float32).to(device)")
s = s.replace("dtype=torch.bfloat16", "dtype=torch.float32")
path.write_text(s)
def done_cells() -> set[tuple[str, str, str]]:
seen = set()
if ROWS.exists():
for line in ROWS.read_text().splitlines():
if not line.strip():
continue
try:
r = json.loads(line)
seen.add((r["model"], r["voice"], r["id"]))
except Exception:
continue
return seen
def sample_latent(student, z, history, noise, n_steps):
y = noise
for i in range(n_steps):
dt = 1.0 / n_steps
v = student(
x=y, t=torch.full((y.shape[0],), i / n_steps, device=y.device),
c=z, latent_history=history, mask=None,
)[:, -1:, :]
y = y + v * dt
return y
def ar_generate(mm, dit, n_steps, prompt, text, prompt_latent, prompt_text, spk_emb, frames):
"""Autoregressive rollout with a chosen flow head. Length is forced to the
teacher's frame count, so this measures content and timbre, not stopping."""
device = next(mm.parameters()).device
dtype = next(mm.parameters()).dtype
with torch.no_grad():
input_ids, inputs_embeds = mm.prepare_input_embed(
prompt=prompt, text=text, spk_emb=spk_emb,
prompt_latent=prompt_latent, prompt_text=prompt_text,
)
attention_mask = torch.ones(input_ids.shape, device=device)
position_ids = (attention_mask.cumsum(-1) - 1).masked_fill_(attention_mask == 0, 1)
past = None
history = torch.zeros(1, mm.history_patch_size, mm.latent_dim, device=device, dtype=dtype)
if prompt_latent is not None:
start = mm.history_patch_size - prompt_latent.size(1)
if start < 0:
history[:] = prompt_latent[:, -start:, :]
else:
history[:, start:, :] = prompt_latent
out = []
for _ in range(frames):
with torch.no_grad():
hidden = mm.model(
attention_mask=attention_mask, position_ids=position_ids,
inputs_embeds=inputs_embeds, audio_mask=None, image_mask=None,
output_hidden_states=True, return_dict=True, use_cache=True,
past_key_values=past,
)
past = hidden.past_key_values
z = hidden.hidden_states[-1][:, -1:, :]
noise = torch.randn(1, 1, mm.latent_dim, device=device, dtype=dtype)
latent = sample_latent(dit, z, history, noise, n_steps).to(device=device, dtype=dtype)
out.append(latent)
inputs_embeds = mm.linear_proj_audio(latent)
position_ids = position_ids[:, -1:] + 1
attention_mask = torch.ones(inputs_embeds.shape[0], 1, device=device)
history[:, :-mm.patch_size, :] = history[:, mm.patch_size:, :].clone()
history[:, -mm.patch_size:, :] = latent
with torch.no_grad():
speech = mm.audio.decode(torch.cat(out, dim=1), past_key_values=None,
use_cache=False, last_chunk=True)[0]
return speech.cpu().float().reshape(-1).numpy()
def load_student(name: str, teacher_dit, device):
from huggingface_hub import hf_hub_download
repo, steps = STUDENT_REPOS[name]
path = hf_hub_download(repo_id=repo, filename=f"distill/student_{steps}step.pt")
state = torch.load(path, map_location="cpu")
state = {k: v.float() for k, v in state.items()}
import copy
dit = copy.deepcopy(teacher_dit)
dit.load_state_dict(state)
return dit.to(device).eval(), steps
def build_asr():
from transformers import pipeline
kwargs = {"model": "openai/whisper-small", "device": 0}
try:
pipe = pipeline("automatic-speech-recognition", torch_dtype=torch.float16, **kwargs)
except Exception:
pipe = pipeline("automatic-speech-recognition", **kwargs)
def asr(wav: np.ndarray, sr: int = 44100) -> str:
# No initial_prompt here: transformers 4.46 rejects it as a generate
# kwarg. Whisper's traditional-character output is handled downstream
# by eval/score.py, which normalises before counting edits.
out = pipe(
{"array": wav.astype(np.float32), "sampling_rate": sr},
generate_kwargs={"language": "zh", "task": "transcribe"},
)
return out.get("text", "") if isinstance(out, dict) else str(out)
return asr
def main() -> None:
torch.manual_seed(SEED)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(SEED)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
log(f"device={device} limit={LIMIT} models={MODELS} voices={VOICE_FILTER or 'all'}")
sys.path.insert(0, str(REPO))
patch_fp32()
from ming_tts import MingTTS
import tts_tasks
sentences = json.loads((DATA / "eval/sentences.json").read_text())["items"][:LIMIT]
voices = json.loads((DATA / "config/ming_voices.json").read_text())["voices"]
if VOICE_FILTER:
voices = [v for v in voices if v["id"] in VOICE_FILTER]
log(f"sentences={len(sentences)} voices={[v['id'] for v in voices]}")
tts = MingTTS(str(MODEL_DIR), device=str(device), text_normalization=False)
mm = tts.model
dtype = next(mm.parameters()).dtype
prompt = tts_tasks.build_tts_prompt("Chinese")
teacher_dit = mm.flowloss.cfm.model.float().eval()
for p in teacher_dit.parameters():
p.requires_grad_(False)
students = {}
for name in MODELS:
if name in STUDENT_REPOS:
students[name], steps = load_student(name, teacher_dit, device)
log(f"loaded student {name} ({steps} steps)")
asr = build_asr()
spk = tts.spkemb_extractor
mel = torchaudio.transforms.MelSpectrogram(
sample_rate=44100, n_fft=1024, hop_length=256, n_mels=80, power=2.0
)
def speaker_cos(ref: np.ndarray, hyp: np.ndarray) -> float:
a = torchaudio.functional.resample(torch.from_numpy(ref).float(), 44100, 16000).unsqueeze(0)
b = torchaudio.functional.resample(torch.from_numpy(hyp).float(), 44100, 16000).unsqueeze(0)
ea, eb = spk(a).reshape(-1).float(), spk(b).reshape(-1).float()
return float(torch.dot(ea, eb) / (ea.norm() * eb.norm() + 1e-9))
def mel_l1(ref: np.ndarray, hyp: np.ndarray) -> float:
mr = torch.log(mel(torch.from_numpy(ref).float().unsqueeze(0)) + 1e-6).squeeze(0)
mh = torch.log(mel(torch.from_numpy(hyp).float().unsqueeze(0)) + 1e-6).squeeze(0)
n = min(mr.shape[-1], mh.shape[-1])
return float((mr[..., :n] - mh[..., :n]).abs().mean())
seen = done_cells()
log(f"resume: {len(seen)} cells already done")
for voice in voices:
vid = voice["id"]
wav_path = DATA / voice["promptWav"]
prompt_text = voice["promptText"]
prompt_latent, spk_emb = tts.prepare_prompt(str(wav_path), use_spk_emb=True)
prompt_latent = prompt_latent.to(device=device, dtype=dtype)
spk_emb = [x.to(device=device, dtype=dtype) for x in spk_emb]
for item in sentences:
sid, text = item["id"], item["text"]
ref_wav_path = WAV / "teacher" / vid / f"{sid}.wav"
if ("teacher", vid, sid) not in seen:
try:
t0 = time.time()
wav = tts.synthesize(
task="TTS", text=text, language_id="Chinese",
prompt_wav_path=str(wav_path), prompt_text=prompt_text,
use_spk_emb=True, raw_text=True,
max_decode_steps=max(16, min(400, int(len(text) * 3.0))),
)
arr = wav.cpu().numpy().reshape(-1)
ref_wav_path.parent.mkdir(parents=True, exist_ok=True)
torchaudio.save(str(ref_wav_path), torch.from_numpy(arr).unsqueeze(0), 44100)
frames = max(16, int(round(arr.size / 3528)))
with open(ROWS, "a") as f:
f.write(json.dumps({
"model": "teacher", "voice": vid, "id": sid, "bucket": item["bucket"],
"text": text, "hyp_text": asr(arr), "frames": frames,
"seconds": arr.size / 44100, "wall": time.time() - t0,
}, ensure_ascii=False) + "\n")
seen.add(("teacher", vid, sid))
log(f"teacher {vid} {sid} {arr.size/44100:.2f}s in {time.time()-t0:.1f}s")
except Exception:
log(f"teacher {vid} {sid} failed\n{traceback.format_exc()}")
continue
if not ref_wav_path.exists():
continue
ref_audio, _ = torchaudio.load(str(ref_wav_path))
ref_audio = ref_audio.reshape(-1).numpy()
frames = max(16, int(round(ref_audio.size / 3528)))
for name, dit in students.items():
if (name, vid, sid) in seen:
continue
try:
t0 = time.time()
torch.manual_seed(SEED)
arr = ar_generate(mm, dit, STUDENT_REPOS[name][1], prompt, text,
prompt_latent, prompt_text, spk_emb, frames)
path = WAV / name / vid / f"{sid}.wav"
path.parent.mkdir(parents=True, exist_ok=True)
torchaudio.save(str(path), torch.from_numpy(arr).unsqueeze(0), 44100)
with open(ROWS, "a") as f:
f.write(json.dumps({
"model": name, "voice": vid, "id": sid, "bucket": item["bucket"],
"text": text, "hyp_text": asr(arr),
"spk_cos": speaker_cos(ref_audio, arr),
"mel_l1": mel_l1(ref_audio, arr),
"dur_ratio": arr.size / max(1, ref_audio.size),
"wall": time.time() - t0,
}, ensure_ascii=False) + "\n")
seen.add((name, vid, sid))
log(f"{name} {vid} {sid} in {time.time()-t0:.1f}s")
except Exception:
log(f"{name} {vid} {sid} failed\n{traceback.format_exc()}")
log("DONE")
if __name__ == "__main__":
try:
main()
except Exception:
log(traceback.format_exc())
raise
|