186 lines
7.7 KiB
Python
186 lines
7.7 KiB
Python
from __future__ import annotations
|
|
|
|
from collections.abc import Sequence
|
|
from typing import TYPE_CHECKING, Any, Optional, Union
|
|
|
|
import numpy as np
|
|
import onnxruntime
|
|
from numpy.typing import NDArray
|
|
|
|
from style_bert_vits2.constants import Languages
|
|
from style_bert_vits2.nlp import bert_models, onnx_bert_models
|
|
from style_bert_vits2.nlp.japanese.g2p import text_to_sep_kata
|
|
from style_bert_vits2.utils import get_onnx_device_options
|
|
|
|
|
|
if TYPE_CHECKING:
|
|
import torch
|
|
|
|
|
|
def extract_bert_feature(
|
|
text: str,
|
|
word2ph: list[int],
|
|
device: str,
|
|
assist_text: Optional[str] = None,
|
|
assist_text_weight: float = 0.7,
|
|
) -> torch.Tensor:
|
|
"""
|
|
日本語のテキストから BERT の特徴量を抽出する (PyTorch 推論)
|
|
|
|
Args:
|
|
text (str): 日本語のテキスト
|
|
word2ph (list[int]): 元のテキストの各文字に音素が何個割り当てられるかを表すリスト
|
|
device (str): 推論に利用するデバイス
|
|
assist_text (Optional[str], optional): 補助テキスト (デフォルト: None)
|
|
assist_text_weight (float, optional): 補助テキストの重み (デフォルト: 0.7)
|
|
|
|
Returns:
|
|
torch.Tensor: BERT の特徴量
|
|
"""
|
|
|
|
import torch
|
|
|
|
# 各単語が何文字かを作る `word2ph` を使う必要があるので、読めない文字は必ず無視する
|
|
# でないと `word2ph` の結果とテキストの文字数結果が整合性が取れない
|
|
text = "".join(text_to_sep_kata(text, raise_yomi_error=False)[0])
|
|
if assist_text:
|
|
assist_text = "".join(text_to_sep_kata(assist_text, raise_yomi_error=False)[0])
|
|
|
|
if device == "cuda" and not torch.cuda.is_available():
|
|
device = "cpu"
|
|
model = bert_models.load_model(Languages.JP, device_map=device)
|
|
bert_models.transfer_model(Languages.JP, device)
|
|
|
|
style_res_mean = None
|
|
with torch.no_grad():
|
|
tokenizer = bert_models.load_tokenizer(Languages.JP)
|
|
inputs = tokenizer(text, return_tensors="pt")
|
|
for i in inputs:
|
|
inputs[i] = inputs[i].to(device) # type: ignore
|
|
res = model(**inputs, output_hidden_states=True)
|
|
res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()
|
|
if assist_text:
|
|
style_inputs = tokenizer(assist_text, return_tensors="pt")
|
|
for i in style_inputs:
|
|
style_inputs[i] = style_inputs[i].to(device) # type: ignore
|
|
style_res = model(**style_inputs, output_hidden_states=True)
|
|
style_res = torch.cat(style_res["hidden_states"][-3:-2], -1)[0].cpu()
|
|
style_res_mean = style_res.mean(0)
|
|
|
|
assert len(word2ph) == len(text) + 2, text
|
|
word2phone = word2ph
|
|
phone_level_feature = []
|
|
for i in range(len(word2phone)):
|
|
if assist_text:
|
|
assert style_res_mean is not None
|
|
repeat_feature = (
|
|
res[i].repeat(word2phone[i], 1) * (1 - assist_text_weight)
|
|
+ style_res_mean.repeat(word2phone[i], 1) * assist_text_weight
|
|
)
|
|
else:
|
|
repeat_feature = res[i].repeat(word2phone[i], 1)
|
|
phone_level_feature.append(repeat_feature)
|
|
|
|
phone_level_feature = torch.cat(phone_level_feature, dim=0)
|
|
|
|
return phone_level_feature.T
|
|
|
|
|
|
def extract_bert_feature_onnx(
|
|
text: str,
|
|
word2ph: list[int],
|
|
onnx_providers: Sequence[Union[str, tuple[str, dict[str, Any]]]],
|
|
assist_text: Optional[str] = None,
|
|
assist_text_weight: float = 0.7,
|
|
) -> NDArray[Any]:
|
|
"""
|
|
日本語のテキストから BERT の特徴量を抽出する (ONNX 推論)
|
|
|
|
Args:
|
|
text (str): 日本語のテキスト
|
|
word2ph (list[int]): 元のテキストの各文字に音素が何個割り当てられるかを表すリスト
|
|
onnx_providers (list[str]): ONNX 推論で利用する ExecutionProvider (CPUExecutionProvider, CUDAExecutionProvider など)
|
|
assist_text (Optional[str], optional): 補助テキスト (デフォルト: None)
|
|
assist_text_weight (float, optional): 補助テキストの重み (デフォルト: 0.7)
|
|
|
|
Returns:
|
|
NDArray[Any]: BERT の特徴量
|
|
"""
|
|
|
|
# 各単語が何文字かを作る `word2ph` を使う必要があるので、読めない文字は必ず無視する
|
|
# でないと `word2ph` の結果とテキストの文字数結果が整合性が取れない
|
|
text = "".join(text_to_sep_kata(text, raise_yomi_error=False)[0])
|
|
if assist_text:
|
|
assist_text = "".join(text_to_sep_kata(assist_text, raise_yomi_error=False)[0])
|
|
|
|
# トークナイザーとモデルの読み込み
|
|
tokenizer = onnx_bert_models.load_tokenizer(Languages.JP)
|
|
session = onnx_bert_models.load_model(
|
|
language=Languages.JP,
|
|
onnx_providers=onnx_providers,
|
|
)
|
|
input_names = [input.name for input in session.get_inputs()]
|
|
output_name = session.get_outputs()[0].name
|
|
|
|
# 入力テンソルの転送に使用するデバイス種別, デバイス ID, 実行オプションを取得
|
|
device_type, device_id, run_options = get_onnx_device_options(session, onnx_providers) # fmt: skip
|
|
|
|
# 入力をテンソルに変換
|
|
inputs = tokenizer(text, return_tensors="np")
|
|
input_tensor = [
|
|
inputs["input_ids"].astype(np.int64), # type: ignore
|
|
inputs["attention_mask"].astype(np.int64), # type: ignore
|
|
]
|
|
# 推論デバイスに入力テンソルを割り当て
|
|
## GPU 推論の場合、device_type + device_id に対応する GPU デバイスに入力テンソルが割り当てられる
|
|
io_binding = session.io_binding()
|
|
for name, value in zip(input_names, input_tensor):
|
|
gpu_tensor = onnxruntime.OrtValue.ortvalue_from_numpy(
|
|
value, device_type, device_id
|
|
)
|
|
io_binding.bind_ortvalue_input(name, gpu_tensor)
|
|
# text から BERT 特徴量を抽出
|
|
io_binding.bind_output(output_name, device_type)
|
|
session.run_with_iobinding(io_binding, run_options=run_options)
|
|
res = io_binding.get_outputs()[0].numpy()
|
|
|
|
style_res_mean = None
|
|
if assist_text:
|
|
# 入力をテンソルに変換
|
|
style_inputs = tokenizer(assist_text, return_tensors="np")
|
|
style_input_tensor = [
|
|
style_inputs["input_ids"].astype(np.int64), # type: ignore
|
|
style_inputs["attention_mask"].astype(np.int64), # type: ignore
|
|
]
|
|
# 推論デバイスに入力テンソルを割り当て
|
|
## GPU 推論の場合、device_type + device_id に対応する GPU デバイスに入力テンソルが割り当てられる
|
|
io_binding = session.io_binding() # IOBinding は作り直す必要がある
|
|
for name, value in zip(input_names, style_input_tensor):
|
|
gpu_tensor = onnxruntime.OrtValue.ortvalue_from_numpy(
|
|
value, device_type, device_id
|
|
)
|
|
io_binding.bind_ortvalue_input(name, gpu_tensor)
|
|
# assist_text から BERT 特徴量を抽出
|
|
io_binding.bind_output(output_name, device_type)
|
|
session.run_with_iobinding(io_binding, run_options=run_options)
|
|
style_res = io_binding.get_outputs()[0].numpy()
|
|
style_res_mean = np.mean(style_res, axis=0)
|
|
|
|
assert len(word2ph) == len(text) + 2, text
|
|
word2phone = word2ph
|
|
phone_level_feature = []
|
|
for i in range(len(word2phone)):
|
|
if assist_text:
|
|
assert style_res_mean is not None
|
|
repeat_feature = (
|
|
np.tile(res[i], (word2phone[i], 1)) * (1 - assist_text_weight)
|
|
+ np.tile(style_res_mean, (word2phone[i], 1)) * assist_text_weight
|
|
)
|
|
else:
|
|
repeat_feature = np.tile(res[i], (word2phone[i], 1))
|
|
phone_level_feature.append(repeat_feature)
|
|
|
|
phone_level_feature = np.concatenate(phone_level_feature, axis=0)
|
|
|
|
return phone_level_feature.T
|