bert_vits2/text/japanese_bert.py

import os

import torch
from transformers import AutoTokenizer, AutoModelForMaskedLM

import config
from logger import logger
from utils.download import download_and_verify
from config import DEVICE as device

URLS = [
    "https://huggingface.co/cl-tohoku/bert-base-japanese-v3/resolve/main/pytorch_model.bin",
]
TARGET_PATH = os.path.join(config.ABS_PATH, "bert_vits2/bert/bert-base-japanese-v3/pytorch_model.bin")
EXPECTED_MD5 = None

if not os.path.exists(TARGET_PATH):
    success, message = download_and_verify(URLS, TARGET_PATH, EXPECTED_MD5)

try:
    logger.info("Loading bert-base-japanese-v3...")
    tokenizer = AutoTokenizer.from_pretrained(config.ABS_PATH + "/bert_vits2/bert/bert-base-japanese-v3")
    model = AutoModelForMaskedLM.from_pretrained(config.ABS_PATH + "/bert_vits2/bert/bert-base-japanese-v3").to(
        device)
    logger.info("Loading finished.")
except Exception as e:
    logger.error(e)
    logger.error(f"Please download pytorch_model.bin from cl-tohoku/bert-base-japanese-v3.")


def get_bert_feature(text, word2ph, device=config.DEVICE):
    with torch.no_grad():
        inputs = tokenizer(text, return_tensors="pt")
        for i in inputs:
            inputs[i] = inputs[i].to(device)
        res = model(**inputs, output_hidden_states=True)
        res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()
    assert inputs["input_ids"].shape[-1] == len(word2ph)
    word2phone = word2ph
    phone_level_feature = []
    for i in range(len(word2phone)):
        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