app.py

import os
from huggingface_hub import snapshot_download
snapshot_download(repo_id="None1145/GPT-SoVITS-Lappland-the-Decadenza", cache_dir="./Models")
snapshot_download(repo_id="None1145/GPT-SoVITS-Base", cache_dir="./PretrainedModels")
cnhubert_base_path = "PretrainedModels/chinese-hubert-base"
bert_path = "PretrainedModels/chinese-roberta-wwm-ext-large"

import gradio as gr
from transformers import AutoModelForMaskedLM, AutoTokenizer
import sys,torch,numpy as np
from pathlib import Path
import os,pdb,utils,librosa,math,traceback,requests,argparse,torch,multiprocessing,pandas as pd,torch.multiprocessing as mp,soundfile
# torch.backends.cuda.sdp_kernel("flash")
# torch.backends.cuda.enable_flash_sdp(True)
# torch.backends.cuda.enable_mem_efficient_sdp(True)  # Not avaliable if torch version is lower than 2.0
# torch.backends.cuda.enable_math_sdp(True)
from random import shuffle
from AR.utils import get_newest_ckpt
from glob import glob
from tqdm import tqdm
from feature_extractor import cnhubert
cnhubert.cnhubert_base_path=cnhubert_base_path
from io import BytesIO
from module.models import SynthesizerTrn
from AR.models.t2s_lightning_module import Text2SemanticLightningModule
from AR.utils.io import load_yaml_config
from text import cleaned_text_to_sequence
from text.cleaner import text_to_sequence, clean_text
from time import time as ttime
from module.mel_processing import spectrogram_torch
from my_utils import load_audio
import re

import logging
logging.getLogger('httpx').setLevel(logging.WARNING)
logging.getLogger('httpcore').setLevel(logging.WARNING)
logging.getLogger('multipart').setLevel(logging.WARNING)

device = "cpu"
is_half = False

tokenizer = AutoTokenizer.from_pretrained(bert_path)
bert_model=AutoModelForMaskedLM.from_pretrained(bert_path)
if(is_half==True):bert_model=bert_model.half().to(device)
else:bert_model=bert_model.to(device)
# bert_model=bert_model.to(device)
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)#####输入是long不用管精度问题，精度随bert_model
        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)
    # if(is_half==True):phone_level_feature=phone_level_feature.half()
    return phone_level_feature.T


def load_model(sovits_path, gpt_path):
    n_semantic = 1024
    dict_s2 = torch.load(sovits_path, map_location="cpu")
    hps = dict_s2["config"]

    class DictToAttrRecursive:
        def __init__(self, input_dict):
            for key, value in input_dict.items():
                if isinstance(value, dict):
                    # 如果值是字典，递归调用构造函数
                    setattr(self, key, DictToAttrRecursive(value))
                else:
                    setattr(self, key, value)

    hps = DictToAttrRecursive(hps)
    hps.model.semantic_frame_rate = "25hz"
    dict_s1 = torch.load(gpt_path, map_location="cpu")
    config = dict_s1["config"]
    ssl_model = cnhubert.get_model()
    if (is_half == True):
        ssl_model = ssl_model.half().to(device)
    else:
        ssl_model = ssl_model.to(device)

    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 (is_half == True):
        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)
    hz = 50
    max_sec = config['data']['max_sec']
    # t2s_model = Text2SemanticLightningModule.load_from_checkpoint(checkpoint_path=gpt_path, config=config, map_location="cpu")#########todo
    t2s_model = Text2SemanticLightningModule(config, "ojbk", is_train=False)
    t2s_model.load_state_dict(dict_s1["weight"])
    if (is_half == True): t2s_model = t2s_model.half()
    t2s_model = t2s_model.to(device)
    t2s_model.eval()
    total = sum([param.nelement() for param in t2s_model.parameters()])
    print("Number of parameter: %.2fM" % (total / 1e6))
    return vq_model, ssl_model, t2s_model, hps, config, hz, max_sec


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 create_tts_fn(vq_model, ssl_model, t2s_model, hps, config, hz, max_sec):
    def tts_fn(ref_wav_path, prompt_text, prompt_language, text, text_language):
        t0 = ttime()
        prompt_text=prompt_text.strip("\n")
        prompt_language,text=prompt_language,text.strip("\n")
        print(text)
        if len(text) > 50:
            return f"Error: Text is too long, ({len(text)}>50)", None
        with torch.no_grad():
            wav16k, sr = librosa.load(ref_wav_path, sr=16000)  # 派蒙
            wav16k = torch.from_numpy(wav16k)
            if(is_half==True):wav16k=wav16k.half().to(device)
            else:wav16k=wav16k.to(device)
            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]
        t1 = ttime()
        phones1, word2ph1, norm_text1 = clean_text(prompt_text, prompt_language)
        phones1=cleaned_text_to_sequence(phones1)
        texts=text.split("\n")
        audio_opt = []
        zero_wav=np.zeros(int(hps.data.sampling_rate*0.3),dtype=np.float16 if is_half==True else np.float32)
        for text in texts:
            phones2, word2ph2, norm_text2 = clean_text(text, text_language)
            phones2 = cleaned_text_to_sequence(phones2)
            if(prompt_language=="zh"):bert1 = get_bert_feature(norm_text1, word2ph1).to(device)
            else:bert1 = torch.zeros((1024, len(phones1)),dtype=torch.float16 if is_half==True else torch.float32).to(device)
            if(text_language=="zh"):bert2 = get_bert_feature(norm_text2, word2ph2).to(device)
            else:bert2 = torch.zeros((1024, len(phones2))).to(bert1)
            bert = torch.cat([bert1, bert2], 1)

            all_phoneme_ids = torch.LongTensor(phones1+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)
            t2 = ttime()
            with torch.no_grad():
                # pred_semantic = t2s_model.model.infer(
                pred_semantic,idx = t2s_model.model.infer_panel(
                    all_phoneme_ids,
                    all_phoneme_len,
                    prompt,
                    bert,
                    # prompt_phone_len=ph_offset,
                    top_k=config['inference']['top_k'],
                    early_stop_num=hz * max_sec)
            t3 = ttime()
            # 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
            audio_opt.append(audio)
            audio_opt.append(zero_wav)
            t4 = ttime()
        print("%.3f\t%.3f\t%.3f\t%.3f" % (t1 - t0, t2 - t1, t3 - t2, t4 - t3))
        return "Success", (hps.data.sampling_rate,(np.concatenate(audio_opt,0)*32768).astype(np.int16))
    return tts_fn


splits={"，","。","？","！",",",".","?","!","~",":","：","—","…",}#不考虑省略号
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 change_reference_audio(prompt_text, transcripts):
    return transcripts[prompt_text]

models = []
models_info = {}
models_folder_path = "./Models/None1145"
folder_names = [name for name in os.listdir(models_folder_path) if os.path.isdir(os.path.join(models_folder_path, name))]
for folder_name in folder_names:
    speaker = folder_name[11:]
    models_info[speaker] = {}
    models_info[speaker]["title"] = speaker
    pattern = re.compile(r"s(\d+)\.pth$")
    max_value = -1
    max_file = None
    sovits_path = f"{models_folder_path}/{folder_name}/SoVITS_weights"
    for filename in os.listdir(sovits_path):
        match = pattern.search(filename)
        if match:
            value = int(match.group(1))
            if value > max_value:
                max_value = value
                max_file = filename
    models_info[speaker]["sovits_weight"] = f"{sovits_path}/{max_file}"
    pattern = re.compile(r"e(\d+)\.ckpt$")
    max_value = -1
    max_file = None
    gpt_path = f"{models_folder_path}/{folder_name}/GPT_weights"
    for filename in os.listdir(gpt_path):
        match = pattern.search(filename)
        if match:
            value = int(match.group(1))
            if value > max_value:
                max_value = value
                max_file = filename
    models_info[speaker]["gpt_weight"] = f"{gpt_path}/{max_file}"
    data_path = f"{models_folder_path}/{folder_name}/Data"
    models_info[speaker]["transcript"] = {}
    with open(f"{data_path}/{speaker}.list", "r", encoding="utf-8") as f:
        for line in f.read().split("\n"):
            wav = f"{models_folder_path}/{folder_name}/Data/{line.split("|")[0].split("/")[1]}"
            text = line.split("|")[3]
            models_info[speaker]["transcript"][text] = wav
            models_info[speaker]["example_reference"] = text
for speaker in models_info:
    speaker_info = models_info[speaker]
    title = speaker_info["title"]
    sovits_weight = speaker_info["sovits_weight"]
    gpt_weight = speaker_info["gpt_weight"]
    vq_model, ssl_model, t2s_model, hps, config, hz, max_sec = load_model(sovits_weight, gpt_weight)
    models.append(
        (
            speaker,
            title,
            speaker_info["transcript"],
            speaker_info["example_reference"],
            create_tts_fn(
                vq_model, ssl_model, t2s_model, hps, config, hz, max_sec
            )
        )
    )

with gr.Blocks() as app:
    with gr.Tabs():
        for (name, title, transcript, example_reference, tts_fn) in models:
            with gr.TabItem(name):
                with gr.Row():
                    gr.Markdown(
                        '<div align="center">'
                        f'<a><strong>{title}</strong></a>'
                        '</div>')
                with gr.Row():
                    with gr.Column():
                        prompt_text = gr.Dropdown(
                            label="Transcript of the Reference Audio",
                            value=example_reference,
                            choices=list(transcript.keys())
                        )
                        inp_ref_audio = gr.Audio(
                            label="Reference Audio",
                            type="filepath",
                            interactive=False,
                            value=transcript[example_reference]
                        )
                        transcripts_state = gr.State(value=transcript)
                        prompt_text.change(
                            fn=change_reference_audio,
                            inputs=[prompt_text, transcripts_state],
                            outputs=[inp_ref_audio]
                        )
                        prompt_language = gr.State(value="zh")
                    with gr.Column():
                        text = gr.Textbox(label="Input Text", value="你好。")
                        text_language = gr.Dropdown(
                            label="Language",
                            choices=["zh", "en", "ja"],
                            value="ja"
                        )
                        inference_button = gr.Button("Generate", variant="primary")
                        om = gr.Textbox(label="Output Message")
                        output = gr.Audio(label="Output Audio")
                        inference_button.click(
                            fn=tts_fn,
                            inputs=[inp_ref_audio, prompt_text, prompt_language, text, text_language],
                            outputs=[om, output]
                        )

app.queue().launch()