fill the content in code file.
Browse filesby mistake file pushed without content.
- mio_codec_wrapper.py +215 -0
mio_codec_wrapper.py
CHANGED
|
@@ -0,0 +1,215 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""mio_codec_wrapper.py
|
| 3 |
+
|
| 4 |
+
Wrapper to run MioCodec in two modes:
|
| 5 |
+
- streaming: connects to a websocket (e.g. vLLM/your LLM) that emits
|
| 6 |
+
token embeddings (JSON messages) and decodes them to audio in near
|
| 7 |
+
real-time using the windowed streamer.
|
| 8 |
+
- offline: loads a JSONL file of embeddings and decodes to a WAV file.
|
| 9 |
+
|
| 10 |
+
Requirements (minimal): torch, numpy, websockets (for streaming), soundfile (for offline write), sounddevice (optional playback).
|
| 11 |
+
|
| 12 |
+
Install quickly without cloning full repo (if upstream exists):
|
| 13 |
+
pip install git+https://github.com/Aratako/MioCodec.git
|
| 14 |
+
|
| 15 |
+
If you already have local helper files (`streaming_istft_head.py`, `windowed_streamer.py`), the wrapper will import them; otherwise it will try to use `miocodec` package's model and swap in the streaming head.
|
| 16 |
+
"""
|
| 17 |
+
from __future__ import annotations
|
| 18 |
+
|
| 19 |
+
import argparse
|
| 20 |
+
import asyncio
|
| 21 |
+
import json
|
| 22 |
+
import sys
|
| 23 |
+
from pathlib import Path
|
| 24 |
+
|
| 25 |
+
import numpy as np
|
| 26 |
+
import torch
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
def build_args():
|
| 30 |
+
p = argparse.ArgumentParser()
|
| 31 |
+
p.add_argument("--mode", choices=("stream", "offline"), default="stream")
|
| 32 |
+
p.add_argument("--uri", default="ws://localhost:8000/stream", help="Websocket URI for streaming mode")
|
| 33 |
+
p.add_argument("--input", help="JSONL file with embeddings for offline mode")
|
| 34 |
+
p.add_argument("--out", default="out.wav", help="Output WAV file for offline mode")
|
| 35 |
+
p.add_argument("--device", default="cpu")
|
| 36 |
+
p.add_argument("--repo-id", default="Aratako/MioCodec-25Hz-24kHz")
|
| 37 |
+
p.add_argument("--play", action="store_true", help="Play decoded audio while streaming")
|
| 38 |
+
p.add_argument("--chunk-size", type=int, default=20)
|
| 39 |
+
p.add_argument("--left", type=int, default=10)
|
| 40 |
+
p.add_argument("--right", type=int, default=10)
|
| 41 |
+
return p.parse_args()
|
| 42 |
+
|
| 43 |
+
|
| 44 |
+
def load_model_and_streamer(repo_id: str, device: str, chunk_size: int, left: int, right: int):
|
| 45 |
+
"""Return (model, streamer, sample_rate). Attempts to import local windowed_streamer first."""
|
| 46 |
+
try:
|
| 47 |
+
# Prefer local helper module if present
|
| 48 |
+
from windowed_streamer import build_windowed_model, MioCodecWindowedStreamer, WindowedStreamConfig
|
| 49 |
+
|
| 50 |
+
model = build_windowed_model(repo_id=repo_id, device=device)
|
| 51 |
+
cfg = model.config
|
| 52 |
+
stream_cfg = WindowedStreamConfig(chunk_size=chunk_size, left_context_tokens=left, right_lookahead_tokens=right)
|
| 53 |
+
# create a default global embedding (zero) sized to model's expected dim
|
| 54 |
+
gdim = getattr(cfg, "global_embedding_dim", getattr(cfg, "wave_decoder_dim", 512))
|
| 55 |
+
global_embedding = torch.zeros(gdim, dtype=torch.float32).to(device)
|
| 56 |
+
streamer = MioCodecWindowedStreamer(model, global_embedding, config=stream_cfg)
|
| 57 |
+
return model, streamer, getattr(cfg, "sample_rate", 24000)
|
| 58 |
+
except Exception:
|
| 59 |
+
# Fallback: import MioCodecModel from installed package and swap head
|
| 60 |
+
from miocodec.model import MioCodecModel
|
| 61 |
+
try:
|
| 62 |
+
from streaming_istft_head import ISTFTHead as StreamingISTFTHead
|
| 63 |
+
except Exception:
|
| 64 |
+
StreamingISTFTHead = None
|
| 65 |
+
|
| 66 |
+
model = MioCodecModel.from_pretrained(repo_id).to(device)
|
| 67 |
+
cfg = model.config
|
| 68 |
+
if StreamingISTFTHead is not None:
|
| 69 |
+
old_head = model.istft_head
|
| 70 |
+
new_head = StreamingISTFTHead(dim=cfg.wave_decoder_dim, n_fft=cfg.n_fft, hop_length=cfg.hop_length, padding=cfg.istft_padding).to(device)
|
| 71 |
+
new_head.out.load_state_dict(old_head.out.state_dict())
|
| 72 |
+
model.istft_head = new_head
|
| 73 |
+
|
| 74 |
+
# create a simple streamer-like helper using WindowedStreamConfig from windowed_streamer if available
|
| 75 |
+
try:
|
| 76 |
+
from windowed_streamer import MioCodecWindowedStreamer, WindowedStreamConfig
|
| 77 |
+
|
| 78 |
+
stream_cfg = WindowedStreamConfig(chunk_size=chunk_size, left_context_tokens=left, right_lookahead_tokens=right)
|
| 79 |
+
gdim = getattr(cfg, "global_embedding_dim", getattr(cfg, "wave_decoder_dim", 512))
|
| 80 |
+
global_embedding = torch.zeros(gdim, dtype=torch.float32).to(device)
|
| 81 |
+
streamer = MioCodecWindowedStreamer(model, global_embedding, config=stream_cfg)
|
| 82 |
+
except Exception:
|
| 83 |
+
# Minimal fallback: use a naive one-shot decoder wrapper for full sequences
|
| 84 |
+
streamer = None
|
| 85 |
+
|
| 86 |
+
return model, streamer, getattr(cfg, "sample_rate", 24000)
|
| 87 |
+
|
| 88 |
+
|
| 89 |
+
async def stream_mode(uri: str, model, streamer, samplerate: int, play: bool, device: str):
|
| 90 |
+
"""Connect to websocket; expect messages with {"embedding": [...]} or {"finished": true}."""
|
| 91 |
+
try:
|
| 92 |
+
import websockets
|
| 93 |
+
except Exception:
|
| 94 |
+
print("Install websockets: pip install websockets", file=sys.stderr)
|
| 95 |
+
return
|
| 96 |
+
|
| 97 |
+
sd = None
|
| 98 |
+
if play:
|
| 99 |
+
try:
|
| 100 |
+
import sounddevice as sd
|
| 101 |
+
except Exception:
|
| 102 |
+
sd = None
|
| 103 |
+
|
| 104 |
+
async with websockets.connect(uri) as ws:
|
| 105 |
+
print("connected", uri)
|
| 106 |
+
async for raw in ws:
|
| 107 |
+
try:
|
| 108 |
+
msg = json.loads(raw)
|
| 109 |
+
except Exception:
|
| 110 |
+
continue
|
| 111 |
+
|
| 112 |
+
if msg.get("finished"):
|
| 113 |
+
if streamer is not None:
|
| 114 |
+
final = streamer.finish()
|
| 115 |
+
if final is not None:
|
| 116 |
+
arr = final.cpu().numpy().astype(np.float32)
|
| 117 |
+
if sd is not None:
|
| 118 |
+
sd.play(arr, samplerate)
|
| 119 |
+
sd.wait()
|
| 120 |
+
else:
|
| 121 |
+
print("final chunk ready (not played)")
|
| 122 |
+
break
|
| 123 |
+
|
| 124 |
+
emb = msg.get("embedding")
|
| 125 |
+
if emb is None:
|
| 126 |
+
continue
|
| 127 |
+
tensor = torch.tensor(emb, dtype=torch.float32).to(device)
|
| 128 |
+
if streamer is not None:
|
| 129 |
+
audio = streamer.push_token(tensor)
|
| 130 |
+
if audio is not None:
|
| 131 |
+
arr = audio.cpu().numpy().astype(np.float32)
|
| 132 |
+
if sd is not None:
|
| 133 |
+
sd.play(arr, samplerate)
|
| 134 |
+
# do not wait to allow continuous streaming
|
| 135 |
+
else:
|
| 136 |
+
sys.stdout.buffer.write(arr.tobytes())
|
| 137 |
+
else:
|
| 138 |
+
# no streamer: buffer and decode once finished
|
| 139 |
+
# simple approach: accumulate embeddings on model directly if model supports method
|
| 140 |
+
if not hasattr(stream_mode, "buffer"):
|
| 141 |
+
stream_mode.buffer = []
|
| 142 |
+
stream_mode.buffer.append(tensor.cpu().numpy())
|
| 143 |
+
|
| 144 |
+
|
| 145 |
+
def write_wav(path: Path, arr: np.ndarray, samplerate: int):
|
| 146 |
+
try:
|
| 147 |
+
import soundfile as sf
|
| 148 |
+
|
| 149 |
+
sf.write(str(path), arr, samplerate)
|
| 150 |
+
except Exception:
|
| 151 |
+
# fallback: write raw float32
|
| 152 |
+
path.write_bytes(arr.astype(np.float32).tobytes())
|
| 153 |
+
|
| 154 |
+
|
| 155 |
+
def offline_mode(input_path: Path, out_path: Path, model, streamer, samplerate: int, device: str):
|
| 156 |
+
"""Read JSONL file of {embedding: [...]} lines and decode offline."""
|
| 157 |
+
embeddings = []
|
| 158 |
+
with input_path.open("r") as f:
|
| 159 |
+
for line in f:
|
| 160 |
+
try:
|
| 161 |
+
m = json.loads(line)
|
| 162 |
+
except Exception:
|
| 163 |
+
continue
|
| 164 |
+
if "embedding" in m:
|
| 165 |
+
embeddings.append(torch.tensor(m["embedding"], dtype=torch.float32).to(device))
|
| 166 |
+
|
| 167 |
+
if streamer is not None:
|
| 168 |
+
# push all and finish
|
| 169 |
+
for emb in embeddings:
|
| 170 |
+
streamer.push_token(emb)
|
| 171 |
+
final = streamer.finish()
|
| 172 |
+
if final is not None:
|
| 173 |
+
arr = final.cpu().numpy().astype(np.float32)
|
| 174 |
+
write_wav(out_path, arr, samplerate)
|
| 175 |
+
print("Wrote", out_path)
|
| 176 |
+
return
|
| 177 |
+
|
| 178 |
+
# fallback one-shot: try to concatenate and run model-specific decode path
|
| 179 |
+
if hasattr(model, "istft_head") and hasattr(model, "wave_prenet"):
|
| 180 |
+
# build a content embedding and run forward once (best-effort)
|
| 181 |
+
all_tokens = torch.stack(embeddings, dim=0).unsqueeze(0).to(device)
|
| 182 |
+
with torch.no_grad():
|
| 183 |
+
local_latent = model.wave_prenet(all_tokens)
|
| 184 |
+
if model.wave_conv_upsample is not None:
|
| 185 |
+
local_latent = model.wave_conv_upsample(local_latent.transpose(1, 2)).transpose(1, 2)
|
| 186 |
+
# approximate frames
|
| 187 |
+
frames = local_latent.shape[1] * 100 # heuristic; prefer streamer for correctness
|
| 188 |
+
local_latent = torch.nn.functional.interpolate(local_latent.transpose(1, 2), size=frames, mode=model.config.wave_interpolation_mode).transpose(1, 2)
|
| 189 |
+
local_latent = model.wave_prior_net(local_latent.transpose(1, 2)).transpose(1, 2)
|
| 190 |
+
local_latent = model.wave_decoder(local_latent, condition=torch.zeros(1, 1, local_latent.shape[-1], device=device))
|
| 191 |
+
local_latent = model.wave_post_net(local_latent.transpose(1, 2)).transpose(1, 2)
|
| 192 |
+
audio = model.istft_head(local_latent).squeeze(0).cpu().numpy().astype(np.float32)
|
| 193 |
+
write_wav(out_path, audio, samplerate)
|
| 194 |
+
print("Wrote", out_path)
|
| 195 |
+
return
|
| 196 |
+
|
| 197 |
+
print("Could not decode offline: no streamer and model lacks decode path.")
|
| 198 |
+
|
| 199 |
+
|
| 200 |
+
def main():
|
| 201 |
+
args = build_args()
|
| 202 |
+
device = args.device
|
| 203 |
+
model, streamer, samplerate = load_model_and_streamer(args.repo_id, device, args.chunk_size, args.left, args.right)
|
| 204 |
+
|
| 205 |
+
if args.mode == "stream":
|
| 206 |
+
asyncio.run(stream_mode(args.uri, model, streamer, samplerate, args.play, device))
|
| 207 |
+
else:
|
| 208 |
+
if not args.input:
|
| 209 |
+
print("--input required for offline mode", file=sys.stderr)
|
| 210 |
+
return
|
| 211 |
+
offline_mode(Path(args.input), Path(args.out), model, streamer, samplerate, device)
|
| 212 |
+
|
| 213 |
+
|
| 214 |
+
if __name__ == "__main__":
|
| 215 |
+
main()
|