inference_core.py

import logging
import re
from transformers import AutoModelForMaskedLM, AutoTokenizer
import numpy as np
import librosa
import torch

def script_method(fn, _rcb=None):
    return fn
def script(obj, optimize=True, _frames_up=0, _rcb=None):
    return obj
import torch.jit
script_method1 = torch.jit.script_method
script1 = torch.jit.script
torch.jit.script_method = script_method
torch.jit.script = script

from module import cnhubert
import LangSegment
from module.models import SynthesizerTrn
from AR.models.t2s_lightning_module import Text2SemanticLightningModule
from text import cleaned_text_to_sequence
from text.cleaner import clean_text
from module.mel_processing import spectrogram_torch
import ffmpeg

logging.getLogger("markdown_it").setLevel(logging.ERROR)
logging.getLogger("urllib3").setLevel(logging.ERROR)
logging.getLogger("httpcore").setLevel(logging.ERROR)
logging.getLogger("httpx").setLevel(logging.ERROR)
logging.getLogger("asyncio").setLevel(logging.ERROR)
logging.getLogger("charset_normalizer").setLevel(logging.ERROR)
logging.getLogger("torchaudio._extension").setLevel(logging.ERROR)

cnhubert_base_path = "pretrained/chinese-hubert-base"
bert_path = "pretrained/chinese-roberta-wwm-ext-large"

is_half = True
if torch.cuda.is_available():
    device = "cuda"
    gpu_name = torch.cuda.get_device_name(0)
    if (
            ("16" in gpu_name and "V100" not in gpu_name.upper())
            or "P40" in gpu_name.upper()
            or "P10" in gpu_name.upper()
            or "1060" in gpu_name
            or "1070" in gpu_name
            or "1080" in gpu_name
    ):
        is_half = False
else:
    device = "cpu"
    is_half = False

tokenizer = AutoTokenizer.from_pretrained(bert_path)
bert_model = AutoModelForMaskedLM.from_pretrained(bert_path)

cnhubert.cnhubert_base_path = cnhubert_base_path
ssl_model = cnhubert.get_model()

if is_half:
    bert_model = bert_model.half().to(device)
    ssl_model = ssl_model.half().to(device)
else:
    bert_model = bert_model.to(device)
    ssl_model = ssl_model.to(device)

splits = {"，", "。", "？", "！", ",", ".", "?", "!", "~", ":", "：", "—", "…", }

dtype = torch.float16 if is_half else torch.float32


def load_audio(file, sr):
    try:
        # https://github.com/openai/whisper/blob/main/whisper/audio.py#L26
        # This launches a subprocess to decode audio while down-mixing and resampling as necessary.
        # Requires the ffmpeg CLI and `ffmpeg-python` package to be installed.
        file = (
            file.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
        )  # 防止小白拷路径头尾带了空格和"和回车
        out, _ = (
            ffmpeg.input(file, threads=0)
            .output("-", format="f32le", acodec="pcm_f32le", ac=1, ar=sr)
            .run(cmd=["ffmpeg", "-nostdin"], capture_stdout=True, capture_stderr=True)
        )
    except Exception as e:
        raise RuntimeError(f"Failed to load audio: {e}")

    return np.frombuffer(out, np.float32).flatten()


def get_bert_feature(text, word2ph):
    with torch.no_grad():
        inputs = tokenizer(text, return_tensors="pt")
        for i in inputs:
            inputs[i] = inputs[i].to(device)
        res = bert_model(**inputs, output_hidden_states=True)
        res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()[1:-1]
    assert len(word2ph) == len(text)
    phone_level_feature = []
    for i in range(len(word2ph)):
        repeat_feature = res[i].repeat(word2ph[i], 1)
        phone_level_feature.append(repeat_feature)
    phone_level_feature = torch.cat(phone_level_feature, dim=0)
    return phone_level_feature.T


class DictToAttrRecursive(dict):
    def __init__(self, input_dict):
        super().__init__(input_dict)
        for key, value in input_dict.items():
            if isinstance(value, dict):
                value = DictToAttrRecursive(value)
            self[key] = value
            setattr(self, key, value)

    def __getattr__(self, item):
        try:
            return self[item]
        except KeyError:
            raise AttributeError(f"Attribute {item} not found")

    def __setattr__(self, key, value):
        if isinstance(value, dict):
            value = DictToAttrRecursive(value)
        super(DictToAttrRecursive, self).__setitem__(key, value)
        super().__setattr__(key, value)

    def __delattr__(self, item):
        try:
            del self[item]
        except KeyError:
            raise AttributeError(f"Attribute {item} not found")


def change_sovits_weights(sovits_path):
    global vq_model, hps
    dict_s2 = torch.load(sovits_path, map_location="cpu")
    hps = dict_s2["config"]
    hps = DictToAttrRecursive(hps)
    hps.model.semantic_frame_rate = "25hz"
    vq_model = SynthesizerTrn(
        hps.data.filter_length // 2 + 1,
        hps.train.segment_size // hps.data.hop_length,
        n_speakers=hps.data.n_speakers,
        **hps.model
    )
    if "pretrained" not in sovits_path:
        del vq_model.enc_q
    if is_half:
        vq_model = vq_model.half().to(device)
    else:
        vq_model = vq_model.to(device)
    vq_model.eval()
    vq_model.load_state_dict(dict_s2["weight"], strict=False)


def change_gpt_weights(gpt_path):
    global hz, max_sec, t2s_model, config
    hz = 50
    dict_s1 = torch.load(gpt_path, map_location="cpu")
    config = dict_s1["config"]
    max_sec = config["data"]["max_sec"]
    t2s_model = Text2SemanticLightningModule(config, "****", is_train=False)
    t2s_model.load_state_dict(dict_s1["weight"])
    if is_half:
        t2s_model = t2s_model.half()
    t2s_model = t2s_model.to(device)
    t2s_model.eval()


def get_spepc(hps, filename):
    audio = load_audio(filename, int(hps.data.sampling_rate))
    audio = torch.FloatTensor(audio)
    audio_norm = audio
    audio_norm = audio_norm.unsqueeze(0)
    spec = spectrogram_torch(
        audio_norm,
        hps.data.filter_length,
        hps.data.sampling_rate,
        hps.data.hop_length,
        hps.data.win_length,
        center=False,
    )
    return spec


def clean_text_inf(text, language):
    phones, word2ph, norm_text = clean_text(text, language)
    phones = cleaned_text_to_sequence(phones)
    return phones, word2ph, norm_text


def get_bert_inf(phones, word2ph, norm_text, language):
    language = language.replace("all_", "")
    if language == "zh":
        bert = get_bert_feature(norm_text, word2ph).to(device)  # .to(dtype)
    else:
        bert = torch.zeros(
            (1024, len(phones)),
            dtype=torch.float16 if is_half else torch.float32,
        ).to(device)

    return bert


def get_first(text):
    pattern = "[" + "".join(re.escape(sep) for sep in splits) + "]"
    text = re.split(pattern, text)[0].strip()
    return text


def get_phones_and_bert(text, language):
    if language in {"en", "all_zh", "all_ja"}:
        language = language.replace("all_", "")
        if language == "en":
            LangSegment.setfilters(["en"])
            formattext = " ".join(tmp["text"] for tmp in LangSegment.getTexts(text))
        else:
            # 因无法区别中日文汉字,以用户输入为准
            formattext = text
        while "  " in formattext:
            formattext = formattext.replace("  ", " ")
        phones, word2ph, norm_text = clean_text_inf(formattext, language)
        if language == "zh":
            bert = get_bert_feature(norm_text, word2ph).to(device)
        else:
            bert = torch.zeros(
                (1024, len(phones)),
                dtype=torch.float16 if is_half else torch.float32,
            ).to(device)
    elif language in {"zh", "ja", "auto"}:
        textlist = []
        langlist = []
        LangSegment.setfilters(["zh", "ja", "en"])
        if language == "auto":
            for tmp in LangSegment.getTexts(text):
                langlist.append(tmp["lang"])
                textlist.append(tmp["text"])
        else:
            for tmp in LangSegment.getTexts(text):
                if tmp["lang"] == "en":
                    langlist.append(tmp["lang"])
                else:
                    # 因无法区别中日文汉字,以用户输入为准
                    langlist.append(language)
                textlist.append(tmp["text"])
        print(textlist)
        print(langlist)
        phones_list = []
        bert_list = []
        norm_text_list = []
        for i in range(len(textlist)):
            lang = langlist[i]
            phones, word2ph, norm_text = clean_text_inf(textlist[i], lang)
            bert = get_bert_inf(phones, word2ph, norm_text, lang)
            phones_list.append(phones)
            norm_text_list.append(norm_text)
            bert_list.append(bert)
        bert = torch.cat(bert_list, dim=1)
        phones = sum(phones_list, [])
        norm_text = ''.join(norm_text_list)

    return phones, bert.to(dtype), norm_text


def merge_short_text_in_array(texts, threshold):
    if (len(texts)) < 2:
        return texts
    result = []
    text = ""
    for ele in texts:
        text += ele
        if len(text) >= threshold:
            result.append(text)
            text = ""
    if len(text) > 0:
        if len(result) == 0:
            result.append(text)
        else:
            result[len(result) - 1] += text
    return result


def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language, how_to_cut=0, top_k=20, top_p=0.6,
                temperature=0.6, ref_free=False):
    if prompt_text is None or len(prompt_text) == 0:
        ref_free = True

    if not ref_free:
        prompt_text = prompt_text.strip("\n")
        if prompt_text[-1] not in splits: prompt_text += "。" if prompt_language != "en" else "."
    text = text.strip("\n")
    if text[0] not in splits and len(get_first(text)) < 4: text = "。" + text if text_language != "en" else "." + text
    zero_wav = np.zeros(
        int(hps.data.sampling_rate * 0.3),
        dtype=np.float16 if is_half else np.float32,
    )
    with torch.no_grad():
        wav16k, sr = librosa.load(ref_wav_path, sr=16000)
        if wav16k.shape[0] > 160000 or wav16k.shape[0] < 48000:
            raise OSError("参考音频在3~10秒范围外，请更换！")
        wav16k = torch.from_numpy(wav16k)
        zero_wav_torch = torch.from_numpy(zero_wav)
        if is_half:
            wav16k = wav16k.half().to(device)
            zero_wav_torch = zero_wav_torch.half().to(device)
        else:
            wav16k = wav16k.to(device)
            zero_wav_torch = zero_wav_torch.to(device)
        wav16k = torch.cat([wav16k, zero_wav_torch])
        ssl_content = ssl_model.model(wav16k.unsqueeze(0))[
            "last_hidden_state"
        ].transpose(
            1, 2
        )  # .float()
        codes = vq_model.extract_latent(ssl_content)

        prompt_semantic = codes[0, 0]

    if how_to_cut == 1:
        text = cut1(text)
    elif how_to_cut == 2:
        text = cut2(text)
    elif how_to_cut == 3:
        text = cut3(text)
    elif how_to_cut == 4:
        text = cut4(text)
    elif how_to_cut == 5:
        text = cut5(text)
    while "\n\n" in text:
        text = text.replace("\n\n", "\n")
    texts = text.split("\n")
    texts = merge_short_text_in_array(texts, 5)
    audio_opt = []
    if not ref_free:
        phones1, bert1, norm_text1 = get_phones_and_bert(prompt_text, prompt_language)

    for text in texts:
        # 解决输入目标文本的空行导致报错的问题
        if len(text.strip()) == 0:
            continue
        if text[-1] not in splits:
            text += "。" if text_language != "en" else "."
        phones2, bert2, norm_text2 = get_phones_and_bert(text, text_language)
        if not ref_free:
            bert = torch.cat([bert1, bert2], 1)
            all_phoneme_ids = torch.LongTensor(phones1 + phones2).to(device).unsqueeze(0)
        else:
            bert = bert2
            all_phoneme_ids = torch.LongTensor(phones2).to(device).unsqueeze(0)

        bert = bert.to(device).unsqueeze(0)
        all_phoneme_len = torch.tensor([all_phoneme_ids.shape[-1]]).to(device)
        prompt = prompt_semantic.unsqueeze(0).to(device)
        with torch.no_grad():
            # pred_semantic = t2s_model.model.infer(
            pred_semantic, idx = t2s_model.model.infer_panel(
                all_phoneme_ids,
                all_phoneme_len,
                None if ref_free else prompt,
                bert,
                # prompt_phone_len=ph_offset,
                top_k=top_k,
                top_p=top_p,
                temperature=temperature,
                early_stop_num=hz * max_sec,
            )
        # print(pred_semantic.shape,idx)
        pred_semantic = pred_semantic[:, -idx:].unsqueeze(
            0
        )  # .unsqueeze(0)#mq要多unsqueeze一次
        refer = get_spepc(hps, ref_wav_path)  # .to(device)
        if is_half == True:
            refer = refer.half().to(device)
        else:
            refer = refer.to(device)
        # audio = vq_model.decode(pred_semantic, all_phoneme_ids, refer).detach().cpu().numpy()[0, 0]
        audio = (
            vq_model.decode(
                pred_semantic, torch.LongTensor(phones2).to(device).unsqueeze(0), refer
            )
            .detach()
            .cpu()
            .numpy()[0, 0]
        )  ###试试重建不带上prompt部分
        max_audio = np.abs(audio).max()  # 简单防止16bit爆音
        if max_audio > 1: audio /= max_audio
        audio_opt.append(audio)
        audio_opt.append(zero_wav)
    yield hps.data.sampling_rate, (np.concatenate(audio_opt, 0) * 32768).astype(
        np.int16
    )


def split(todo_text):
    todo_text = todo_text.replace("……", "。").replace("——", "，")
    if todo_text[-1] not in splits:
        todo_text += "。"
    i_split_head = i_split_tail = 0
    len_text = len(todo_text)
    todo_texts = []
    while 1:
        if i_split_head >= len_text:
            break  # 结尾一定有标点，所以直接跳出即可，最后一段在上次已加入
        if todo_text[i_split_head] in splits:
            i_split_head += 1
            todo_texts.append(todo_text[i_split_tail:i_split_head])
            i_split_tail = i_split_head
        else:
            i_split_head += 1
    return todo_texts


def cut1(inp):
    inp = inp.strip("\n")
    inps = split(inp)
    split_idx = list(range(0, len(inps), 4))
    split_idx[-1] = None
    if len(split_idx) > 1:
        opts = []
        for idx in range(len(split_idx) - 1):
            opts.append("".join(inps[split_idx[idx]: split_idx[idx + 1]]))
    else:
        opts = [inp]
    return "\n".join(opts)


def cut2(inp):
    inp = inp.strip("\n")
    inps = split(inp)
    if len(inps) < 2:
        return inp
    opts = []
    summ = 0
    tmp_str = ""
    for i in range(len(inps)):
        summ += len(inps[i])
        tmp_str += inps[i]
        if summ > 50:
            summ = 0
            opts.append(tmp_str)
            tmp_str = ""
    if tmp_str != "":
        opts.append(tmp_str)
    # print(opts)
    if len(opts) > 1 and len(opts[-1]) < 50:  ##如果最后一个太短了，和前一个合一起
        opts[-2] = opts[-2] + opts[-1]
        opts = opts[:-1]
    return "\n".join(opts)


def cut3(inp):
    inp = inp.strip("\n")
    return "\n".join(["%s" % item for item in inp.strip("。").split("。")])


def cut4(inp):
    inp = inp.strip("\n")
    return "\n".join(["%s" % item for item in inp.strip(".").split(".")])


# contributed by https://github.com/AI-Hobbyist/GPT-SoVITS/blob/main/GPT_SoVITS/inference_webui.py
def cut5(inp):
    # if not re.search(r'[^\w\s]', inp[-1]):
    # inp += '。'
    inp = inp.strip("\n")
    punds = r'[,.;?!、，。？！;：…]'
    items = re.split(f'({punds})', inp)
    mergeitems = ["".join(group) for group in zip(items[::2], items[1::2])]
    # 在句子不存在符号或句尾无符号的时候保证文本完整
    if len(items) % 2 == 1:
        mergeitems.append(items[-1])
    opt = "\n".join(mergeitems)
    return opt