vall-e/vall_e/inference.py

535 lines
16 KiB
Python
Executable File

import torch
import torchaudio
import soundfile
import time
import logging
import numpy as np
_logger = logging.getLogger(__name__)
from torch import Tensor
from einops import rearrange
from pathlib import Path
from tqdm import tqdm, trange
from .emb import g2p, qnt
from .emb.qnt import trim, trim_random, unload_model, repeat_extend_audio
from .emb.transcribe import transcribe
from .utils import to_device, set_seed, clamp, wrapper as ml
from .config import cfg, Config
from .models import get_models
from .models.lora import enable_lora
from .engines import load_engines, deepspeed_available
from .data import get_phone_symmap, get_lang_symmap, tokenize, sentence_split
from .models import download_model, DEFAULT_MODEL_PATH
if deepspeed_available:
import deepspeed
class TTS():
def __init__( self, config=None, lora=None, device=None, amp=None, dtype=None, attention=None ):
self.loading = True
# yes I can just grab **kwargs and forward them here
self.load_config( config=config, lora=lora, device=device, amp=amp, dtype=dtype, attention=attention )
self.load_model()
self.loading = False
def load_config( self, config=None, lora=None, device=None, amp=None, dtype=None, attention=None ):
if not config:
download_model()
config = DEFAULT_MODEL_PATH
if config.suffix == ".yaml":
_logger.info(f"Loading YAML: {config}")
cfg.load_yaml( config )
elif config.suffix == ".sft":
_logger.info(f"Loading model: {config}")
cfg.load_model( config, lora )
else:
raise Exception(f"Unknown config passed: {config}")
cfg.format( training=False )
cfg.dataset.use_hdf5 = False # could use cfg.load_hdf5(), but why would it ever need to be loaded for inferencing
if amp is None:
amp = cfg.inference.amp
if dtype is None or dtype == "auto":
dtype = cfg.inference.weight_dtype
if device is None:
device = cfg.device
cfg.device = device
cfg.mode = "inferencing"
cfg.trainer.backend = cfg.inference.backend
cfg.trainer.weight_dtype = dtype
cfg.inference.weight_dtype = dtype
self.device = device
self.dtype = cfg.inference.dtype
self.amp = amp
self.batch_size = cfg.inference.batch_size
self.model_kwargs = {}
if attention:
self.model_kwargs["attention"] = attention
def load_model( self ):
load_engines.cache_clear()
unload_model()
self.engines = load_engines(training=False, **self.model_kwargs)
for name, engine in self.engines.items():
if self.dtype != torch.int8:
engine.to(self.device, dtype=self.dtype if not self.amp else torch.float32)
self.engines.eval()
self.symmap = get_phone_symmap()
_logger.info("Loaded model")
def enable_lora( self, enabled=True ):
for name, engine in self.engines.items():
enable_lora( engine.module, mode = enabled )
def disable_lora( self ):
return self.enable_lora( enabled=False )
def encode_text( self, text, language="auto", precheck=True ):
# already a tensor, return it
if isinstance( text, Tensor ):
return text
# check if tokenizes without any unks (for example, if already phonemized text is passes)
if precheck and "<unk>" in self.symmap:
tokens = tokenize( text )
if self.symmap["<unk>"] not in tokens:
return torch.tensor( tokens )
content = g2p.encode(text, language=language)
tokens = tokenize( content )
return torch.tensor( tokens )
def encode_lang( self, language ):
symmap = get_lang_symmap()
id = 0
if language in symmap:
id = symmap[language]
return torch.tensor([ id ])
# to-do: trim before quantizing, instead of after
def encode_audio( self, paths, trim_length=0.0 ):
# already a tensor, return it
if isinstance( paths, Tensor ):
return paths
# split string into paths
if isinstance( paths, str ):
paths = [ Path(p) for p in paths.split(";") ]
# not already a list
if isinstance( paths, Path ):
paths = [ paths ]
proms = []
# merge inputs
for path in paths:
prom = qnt.encode_from_file(path)
if hasattr( prom, "codes" ):
prom = prom.codes
prom = prom[0][:, :].t().to(torch.int16)
proms.append( prom )
res = torch.cat(proms)
if trim_length:
res = repeat_extend_audio( res, int( cfg.dataset.frames_per_second * trim_length ) )
#res = trim( res, int( cfg.dataset.frames_per_second * trim_length ) )
return res
@torch.inference_mode()
def text_embedding( self, input, prom=False ):
model = None
for name, engine in self.engines.items():
model = engine.module
break
if isinstance( input, str ):
input = cfg.tokenizer.encode(input)
if isinstance( input, list ):
input = torch.tensor( input, dtype=torch.uint8, device=self.device )
return model.text_emb( input )
@torch.inference_mode()
def audio_embedding( self, input, prom=False ):
model = None
for name, engine in self.engines.items():
model = engine.module
break
# im really not sure which way is the better way, since the proms_emb and resps_emb have different properties.......
if prom:
return model.proms_emb(
input,
quant_level=input.shape[-1] - 1,
offset=0,
sums=True,
)
return sum([ model.resps_emb(
input[:, :l+1],
offset = 0 if l == 0 else 1, # or maybe set to 1
quant_level = l,
sums = False
) for l in range( input.shape[-1] - 1 ) ])
def modality( self, modality ):
# cringe to handle the best default mode for a given model
if modality == "auto" and cfg.model.name in ["ar+nar", "nar-len"]:
modality = cfg.model.name
return modality
# makes use of being able to batch inputs seamlessly by automatically batching
# this is NOT the default because it absolutely cannot make use of rolling context / prefixing
@torch.inference_mode()
def batched_inference(
self,
texts,
references=None,
languages=None,
text_languages=None,
out_paths=None,
**sampling_kwargs,
):
batch_size = sampling_kwargs.pop("batch_size", self.batch_size)
input_prompt_length = sampling_kwargs.pop("input_prompt_length", 0)
modality = sampling_kwargs.pop("modality", "auto")
seed = sampling_kwargs.pop("seed", None)
use_tqdm = sampling_kwargs.pop("tqdm", True)
use_lora = sampling_kwargs.pop("use_lora", None)
dtype = sampling_kwargs.pop("dtype", self.dtype)
amp = sampling_kwargs.pop("amp", self.amp)
if batch_size < 1:
batch_size = 1
model_ar = None
model_len = None
model_nar = None
for name, engine in self.engines.items():
if model_ar is None and "ar" in engine.hyper_config.capabilities:
model_ar = engine.module
if model_len is None and "len" in engine.hyper_config.capabilities:
model_len = engine.module
if model_nar is None and "nar" in engine.hyper_config.capabilities:
model_nar = engine.module
modality = self.modality( modality )
# force AR+NAR
if modality == "ar+nar":
model_len = None
# force NAR-len
elif modality == "nar-len":
model_ar = None
samples = len(texts)
# fill with null input proms
if not references:
references = [ None for _ in range(samples) ]
# fill with english
if not languages:
languages = [ "auto" for _ in range(samples) ]
if not out_paths:
out_paths = [ None for _ in range(samples) ]
# use the audio language to phonemize the text
if not text_languages:
text_languages = languages
inputs = []
# tensorfy inputs
for i in trange( samples, desc="Preparing batches" ):
# detect language
if languages[i] == "auto":
languages[i] = g2p.detect_language( texts[i] )
texts[i] = self.encode_text( texts[i], language=text_languages[i] )
references[i] = self.encode_audio( references[i], trim_length=input_prompt_length ) if references[i] else None
languages[i] = self.encode_lang( languages[i] )
texts[i] = to_device(texts[i], device=self.device, dtype=torch.uint8 if len(self.symmap) < 256 else torch.int16)
references[i] = to_device(references[i], device=self.device, dtype=torch.int16)
languages[i] = to_device(languages[i], device=self.device, dtype=torch.uint8)
seq_len = texts[i].shape[0] + 1 + (references[i].shape[0] if references[i] is not None else 0) + 1
inputs.append((texts[i], references[i], languages[i], out_paths[i], seq_len))
# attempt to reduce padding
inputs.sort(key=lambda x: x[-1])
# create batches
batches = []
buffer = ([], [], [], [])
for batch in inputs:
# flush
if len(buffer[0]) >= batch_size:
batches.append(buffer)
buffer = ([], [], [], [])
# insert into buffer
for i, x in enumerate( batch[:-1] ):
buffer[i].append(x)
# flush
if buffer:
batches.append(buffer)
buffer = ([], [], [], [])
wavs = []
for texts, proms, langs, out_paths in tqdm(batches, desc="Processing batch"):
seed = set_seed(seed)
batch_size = len(texts)
input_kwargs = dict(
text_list=texts,
proms_list=proms,
lang_list=langs,
disable_tqdm=not use_tqdm,
use_lora=use_lora,
)
with torch.autocast(self.device, dtype=dtype, enabled=amp):
if model_len is not None:
# extra kwargs
duration_padding = sampling_kwargs.pop("duration_padding", 1.05)
len_list = model_len( **input_kwargs, task_list=["len"]*batch_size, **{"max_duration": 5} ) # "max_duration" is max tokens
# add an additional X seconds
len_list = [ int(l * duration_padding) for l in len_list ]
resps_list = model_nar( **input_kwargs, len_list=len_list, task_list=["tts"]*batch_size,
**sampling_kwargs,
)
elif model_ar is not None:
resps_list = model_ar(
**input_kwargs, task_list=["tts"]*batch_size,
**sampling_kwargs,
)
resps_list = model_nar(
**input_kwargs, resps_list=resps_list, task_list=["tts"]*batch_size,
**sampling_kwargs,
)
else:
raise Exception("!")
for resp, out_path in zip( resps_list, out_paths ):
if out_path:
wav, sr = qnt.decode_to_file(resp, out_path, device=self.device)
else:
wav, sr = qnt.decode(resp, device=self.device)
wavs.append(wav)
return wavs
# naive serial inferencing
# will automatically split a text into pieces (if requested) piece by piece
@torch.inference_mode()
def inference(
self,
text,
references,
language="auto",
text_language=None,
task="tts",
out_path=None,
**sampling_kwargs,
):
input_prompt_length = sampling_kwargs.pop("input_prompt_length", 0)
modality = sampling_kwargs.pop("modality", "auto")
seed = sampling_kwargs.pop("seed", None)
use_tqdm = sampling_kwargs.pop("tqdm", True)
use_lora = sampling_kwargs.pop("use_lora", None)
dtype = sampling_kwargs.pop("dtype", self.dtype)
amp = sampling_kwargs.pop("amp", self.amp)
duration_padding = sampling_kwargs.pop("duration_padding", 1.05)
voice_convert = sampling_kwargs.pop("voice_convert", None)
# explicitly require this
if task != "vc":
voice_convert = None
elif voice_convert == None:
raise Exception("Voice conversion requested, but no reference clip provided.")
# transcribe from audio to voice convert from
if voice_convert is not None and not text:
text = transcribe( voice_convert, model_name="openai/whisper-base", align=False )["text"]
lines = sentence_split(text, split_by=sampling_kwargs.get("split_text_by", "sentences"))
wavs = []
sr = None
model_ar = None
model_len = None
model_nar = None
for name, engine in self.engines.items():
if model_ar is None and "ar" in engine.hyper_config.capabilities:
model_ar = engine.module
if model_len is None and "len" in engine.hyper_config.capabilities:
model_len = engine.module
if model_nar is None and "nar" in engine.hyper_config.capabilities:
model_nar = engine.module
seed = set_seed(seed)
modality = self.modality( modality )
# force AR+NAR
if modality == "ar+nar":
model_len = None
# force NAR-len
elif modality == "nar-len":
model_ar = None
if task == "stt":
resp = self.encode_audio( references )
lang = self.encode_lang( language )
resp = to_device(resp, device=self.device, dtype=torch.int16)
lang = to_device(lang, device=self.device, dtype=torch.uint8)
with torch.autocast(self.device, dtype=dtype, enabled=amp):
model = model_ar if model_ar is not None else model_nar
if model is not None:
text_list = model(
text_list=None, proms_list=[resp], lang_list=[lang], resps_list=[resp], task_list=["stt"],
disable_tqdm=not use_tqdm,
use_lora=use_lora,
**sampling_kwargs,
)
else:
raise Exception("!")
text_list = [ cfg.tokenizer.decode( text ).replace(" ", "_").replace(" ", "").replace("_", " ") for text in text_list ]
return text_list[0]
# stuff for rolling context
prefix_context = None
prefix_contexts = []
context_history = sampling_kwargs.get("context_history", 0)
auto_lang = not language or language == "auto"
auto_text_lang = not text_language or text_language == "auto"
vc_utterance = self.encode_audio( voice_convert, trim_length=0 ) if voice_convert else None
prom = self.encode_audio( references, trim_length=input_prompt_length ) if references else None
lang = self.encode_lang( language )
if task in ["ns, sr"]:
prom = [
task,
prom
]
prom = to_device(prom, device=self.device, dtype=torch.int16)
lang = to_device(lang, device=self.device, dtype=torch.uint8)
for line in lines:
if out_path is None:
output_dir = Path("./data/results/")
if not output_dir.exists():
output_dir.mkdir(parents=True, exist_ok=True)
out_path = output_dir / f"{time.time()}.wav"
deduced_language = g2p.detect_language( line ) if auto_lang or auto_text_lang else language
if auto_lang:
language = deduced_language
if auto_text_lang:
text_language = deduced_language
phns = self.encode_text( line, language=text_language )
phns = to_device(phns, device=self.device, dtype=torch.uint8 if len(self.symmap) < 256 else torch.int16)
with torch.autocast(self.device, dtype=dtype, enabled=amp):
input_kwargs = dict(
text_list=[phns],
proms_list=[prom],
lang_list=[lang],
disable_tqdm=not use_tqdm,
use_lora=use_lora,
)
if model_len is not None:
# skip calculating len_list if possible
if task in ["ns, sr"]:
len_list = [ prom[1].shape[0] ]
elif vc_utterance is not None:
len_list = [ vc_utterance.shape[0] ]
else:
len_list = model_len( **input_kwargs, task_list=["len"], **{"max_duration": 5} ) # "max_duration" is max tokens
# add an additional X seconds
len_list = [ int(l * duration_padding) for l in len_list ]
kwargs = {}
if prefix_context is not None:
kwargs["prefix_context"] = prefix_context
if vc_utterance is not None:
kwargs["vc_list"] = [ vc_utterance ]
resps_list = model_nar( **input_kwargs, len_list=len_list, task_list=["tts"],
**(sampling_kwargs | kwargs),
)
elif model_ar is not None:
kwargs = {}
if prefix_context is not None:
kwargs["prefix_context"] = prefix_context
resps_list = model_ar(
**input_kwargs, task_list=["tts"],
**(sampling_kwargs | kwargs),
)
resps_list = model_nar(
**input_kwargs, resps_list=resps_list, task_list=["tts"],
**sampling_kwargs,
)
else:
raise Exception("!")
# to-do: care about batching later
resps = resps_list[0]
# store current context to use as the initial input for later
if context_history > 0:
# add to history
prefix_contexts.append(( phns, resps, resps.shape[0] ))
# then generate the prefix based on how much history to provide
prefix_context = (
[ torch.concat( [ x[0] for x in prefix_contexts[-context_history:] ] ) ],
[ torch.concat( [ x[1] for x in prefix_contexts[-context_history:] ] ) ],
[ sum([ x[2] for x in prefix_contexts[-context_history:] ]) ]
)
# write to file
wav, sr = qnt.decode_to_file(resps, out_path, device=self.device)
# add utterances
wavs.append(wav)
# combine all utterances
return (torch.concat(wavs, dim=-1), sr)