From 4aceab2a6645f1c2e34a9d002e9b1d439603d996 Mon Sep 17 00:00:00 2001 From: KakaruHayate Date: Mon, 22 Jan 2024 20:41:31 +0800 Subject: [PATCH 01/10] =?UTF-8?q?=E6=B7=BB=E5=8A=A0=E6=8A=BD=E5=8D=A1?= =?UTF-8?q?=E5=8A=9F=E8=83=BD?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 添加抽卡功能，并且可以读取ColorSplitter生成的CSV文件进行局部抽取https://github.com/KakaruHayate/ColorSplitter --- GPT_SoVITS/inference_webui.py | 33 +++++++++++++++++++++++++++++++++ 1 file changed, 33 insertions(+) diff --git a/GPT_SoVITS/inference_webui.py b/GPT_SoVITS/inference_webui.py index 246748ae..3f93352b 100644 --- a/GPT_SoVITS/inference_webui.py +++ b/GPT_SoVITS/inference_webui.py @@ -32,6 +32,8 @@ from module.mel_processing import spectrogram_torch from my_utils import load_audio from tools.i18n.i18n import I18nAuto +import random +import pandas as pd i18n = I18nAuto() device = "cuda" @@ -326,6 +328,28 @@ def cut3(inp): return "\n".join(["%s。" % item for item in inp.strip("。").split("。")]) +def select_random_file(folder_path): + files = [f for f in os.listdir(folder_path) if os.path.isfile(os.path.join(folder_path, f))] + return os.path.join(folder_path, random.choice(files)) + + +def get_labels(csv_path): + if csv_path: + df = pd.read_csv(csv_path) + return {"choices": df['clust'].unique().tolist(), "__type__": "update"} + else: + return {"choices": [], "__type__": "update"} + + +def select_random_file_from_csv(csv_path, label, folder_path): + if csv_path: + df = pd.read_csv(csv_path) + files = df[df['clust'] == label]['filename'].apply(os.path.basename).tolist() + return os.path.join(folder_path, random.choice(files)) + else: + return select_random_file(folder_path) + + with gr.Blocks(title="GPT-SoVITS WebUI") as app: gr.Markdown( value=i18n("本软件以MIT协议开源, 作者不对软件具备任何控制力, 使用软件者、传播软件导出的声音者自负全责.
如不认可该条款, 则不能使用或引用软件包内任何代码和文件. 详见根目录LICENSE.") @@ -333,9 +357,18 @@ def cut3(inp): # with gr.Tabs(): # with gr.TabItem(i18n("伴奏人声分离&去混响&去回声")): with gr.Group(): + gr.Markdown(value="随机抽取参考音频，可以导入csv文件限定抽取范围，csv第一列为filename，第二列为clust（类别标签）") + with gr.Row(): + folder_path = gr.Textbox(label="请输入音频文件夹路径", value="") + csv_path = gr.Textbox(label="请输入CSV文件路径（可选）", value="") + label = gr.Dropdown(label="请选择一个标签（可选）", choices=[], value="") + update_labels_button = gr.Button("更新标签", variant="primary") + update_labels_button.click(get_labels, [csv_path], [label]) + random_file_button = gr.Button("随机选择音频文件", variant="primary") gr.Markdown(value=i18n("*请上传并填写参考信息")) with gr.Row(): inp_ref = gr.Audio(label=i18n("请上传参考音频"), type="filepath") + random_file_button.click(select_random_file_from_csv, [csv_path, label, folder_path], [inp_ref]) prompt_text = gr.Textbox(label=i18n("参考音频的文本"), value="") prompt_language = gr.Dropdown( label=i18n("参考音频的语种"),choices=[i18n("中文"),i18n("英文"),i18n("日文")],value=i18n("中文") From ff610bc3f9399804d2758cbfbfab01fc6a7ee59a Mon Sep 17 00:00:00 2001 From: KakaruHayate Date: Thu, 21 Mar 2024 11:19:31 +0800 Subject: [PATCH 02/10] Revert "Merge pull request #821 from KamioRinn/Optimize-English-G2P" This reverts commit 7bc0836d9933402215ce529da7f50bd1a8c63f7f, reversing changes made to b451372316b2611c8607258982d22d9936cde696. --- GPT_SoVITS/text/english.py | 107 ++++++++----------------------------- 1 file changed, 22 insertions(+), 85 deletions(-) diff --git a/GPT_SoVITS/text/english.py b/GPT_SoVITS/text/english.py index 077d33c6..09e20bdd 100644 --- a/GPT_SoVITS/text/english.py +++ b/GPT_SoVITS/text/english.py @@ -8,13 +8,6 @@ from text import symbols -import unicodedata -from builtins import str as unicode -from g2p_en.expand import normalize_numbers -from nltk.tokenize import TweetTokenizer -word_tokenize = TweetTokenizer().tokenize -from nltk import pos_tag - current_file_path = os.path.dirname(__file__) CMU_DICT_PATH = os.path.join(current_file_path, "cmudict.rep") CMU_DICT_FAST_PATH = os.path.join(current_file_path, "cmudict-fast.rep") @@ -195,6 +188,9 @@ def get_dict(): return g2p_dict +eng_dict = get_dict() + + def text_normalize(text): # todo: eng text normalize # 适配中文及 g2p_en 标点 @@ -208,16 +204,6 @@ def text_normalize(text): for p, r in rep_map.items(): text = re.sub(p, r, text) - # 来自 g2p_en 文本格式化处理 - # 增加大写兼容 - text = unicode(text) - text = normalize_numbers(text) - text = ''.join(char for char in unicodedata.normalize('NFD', text) - if unicodedata.category(char) != 'Mn') # Strip accents - text = re.sub("[^ A-Za-z'.,?!\-]", "", text) - text = re.sub(r"(?i)i\.e\.", "that is", text) - text = re.sub(r"(?i)e\.g\.", "for example", text) - return text @@ -234,85 +220,30 @@ def __init__(self): for word in ["AE", "AI", "AR", "IOS", "HUD", "OS"]: del self.cmu[word.lower()] + # "A" 落单不读 "AH0" 读 "EY1" + self.cmu['a'] = [['EY1']] + - def __call__(self, text): - # tokenization - words = word_tokenize(text) - tokens = pos_tag(words) # tuples of (word, tag) - - # steps - prons = [] - for o_word, pos in tokens: - # 还原 g2p_en 小写操作逻辑 - word = o_word.lower() - - if re.search("[a-z]", word) is None: - pron = [word] - # 先把单字母推出去 - elif len(word) == 1: - # 单读 A 发音修正, 这里需要原格式 o_word 判断大写 - if o_word == "A": - pron = ['EY1'] - else: - pron = self.cmu[word][0] - # g2p_en 原版多音字处理 - elif word in self.homograph2features: # Check homograph - pron1, pron2, pos1 = self.homograph2features[word] - if pos.startswith(pos1): - pron = pron1 - else: - pron = pron2 - else: - # 递归查找预测 - pron = self.qryword(word) - - prons.extend(pron) - prons.extend([" "]) - - return prons[:-1] - - - def qryword(self, word): - # 查字典, 单字母除外 - if len(word) > 1 and word in self.cmu: # lookup CMU dict - return self.cmu[word][0] - - # oov 长度小于等于 3 直接读字母 + def predict(self, word): + # 小写 oov 长度小于等于 3 直接读字母 if (len(word) <= 3): - phones = [] - for w in word: - # 单读 A 发音修正, 此处不存在大写的情况 - if w == "a": - phones.extend(['EY1']) - else: - phones.extend(self.cmu[w][0]) - return phones + return [phone for w in word for phone in self(w)] # 尝试分离所有格 if re.match(r"^([a-z]+)('s)$", word): - phones = self.qryword(word[:-2]) - # P T K F TH HH 无声辅音结尾 's 发 ['S'] - if phones[-1] in ['P', 'T', 'K', 'F', 'TH', 'HH']: - phones.extend(['S']) - # S Z SH ZH CH JH 擦声结尾 's 发 ['IH1', 'Z'] 或 ['AH0', 'Z'] - elif phones[-1] in ['S', 'Z', 'SH', 'ZH', 'CH', 'JH']: - phones.extend(['AH0', 'Z']) - # B D G DH V M N NG L R W Y 有声辅音结尾 's 发 ['Z'] - # AH0 AH1 AH2 EY0 EY1 EY2 AE0 AE1 AE2 EH0 EH1 EH2 OW0 OW1 OW2 UH0 UH1 UH2 IY0 IY1 IY2 AA0 AA1 AA2 AO0 AO1 AO2 - # ER ER0 ER1 ER2 UW0 UW1 UW2 AY0 AY1 AY2 AW0 AW1 AW2 OY0 OY1 OY2 IH IH0 IH1 IH2 元音结尾 's 发 ['Z'] - else: - phones.extend(['Z']) - return phones + phone = self(word[:-2]) + phone.extend(['Z']) + return phone # 尝试进行分词，应对复合词 comps = wordsegment.segment(word.lower()) # 无法分词的送回去预测 if len(comps)==1: - return self.predict(word) + return super().predict(word) # 可以分词的递归处理 - return [phone for comp in comps for phone in self.qryword(comp)] + return [phone for comp in comps for phone in self(comp)] _g2p = en_G2p() @@ -327,6 +258,12 @@ def g2p(text): if __name__ == "__main__": + # print(get_dict()) print(g2p("hello")) - print(g2p(text_normalize("e.g. I used openai's AI tool to draw a picture."))) - print(g2p(text_normalize("In this; paper, we propose 1 DSPGAN, a GAN-based universal vocoder."))) \ No newline at end of file + print(g2p("In this; paper, we propose 1 DSPGAN, a GAN-based universal vocoder.")) + # all_phones = set() + # for k, syllables in eng_dict.items(): + # for group in syllables: + # for ph in group: + # all_phones.add(ph) + # print(all_phones) From 5e0ab1839739a04ddd6bae0929adcd8475b8ba80 Mon Sep 17 00:00:00 2001 From: KakaruHayate Date: Sat, 30 Mar 2024 21:33:25 +0800 Subject: [PATCH 03/10] Update webui.py --- webui.py | 16 ++++++++++++++++ 1 file changed, 16 insertions(+) diff --git a/webui.py b/webui.py index e1c36e1e..3a6f2109 100644 --- a/webui.py +++ b/webui.py @@ -79,6 +79,22 @@ # gpu_infos.append("%s\t%s" % ("0", "Apple GPU")) # mem.append(psutil.virtual_memory().total/ 1024 / 1024 / 1024) # 实测使用系统内存作为显存不会爆显存 +# 判断是否有摩尔线程显卡可用 +try: + import torch_musa + use_torch_musa = True +except ImportError: + use_torch_musa = False +if use_torch_musa: + ngpu = torch.musa.device_count() + if torch.musa.is_available(): + for i in range(ngpu): + if_gpu_ok = True + gpu_name = torch.musa.get_device_name(i) + gpu_infos.append("%s\t%s" % ("0", gpu_name)) + mem.append(int(torch.musa.get_device_properties(i).total_memory/ 1024/ 1024/ 1024+ 0.4)) + print("GPT-SoVITS running on MUSA!") + if if_gpu_ok and len(gpu_infos) > 0: gpu_info = "\n".join(gpu_infos) default_batch_size = min(mem) // 2 From 158f85c74521c53e4af14c7f5d5e216c1367d5cc Mon Sep 17 00:00:00 2001 From: KakaruHayate Date: Sat, 30 Mar 2024 21:41:51 +0800 Subject: [PATCH 04/10] Update inference_webui.py --- GPT_SoVITS/inference_webui.py | 436 ++++++++++++++++++++-------------- 1 file changed, 264 insertions(+), 172 deletions(-) diff --git a/GPT_SoVITS/inference_webui.py b/GPT_SoVITS/inference_webui.py index caedc564..237d1b50 100644 --- a/GPT_SoVITS/inference_webui.py +++ b/GPT_SoVITS/inference_webui.py @@ -1,25 +1,47 @@ -import os,re,logging +''' +按中英混合识别 +按日英混合识别 +多语种启动切分识别语种 +全部按中文识别 +全部按英文识别 +全部按日文识别 +''' +import os, re, logging +import LangSegment 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) import pdb +import torch + +device = "cpu" + +try: + import torch_musa + use_torch_musa = True +except ImportError: + use_torch_musa = False +if use_torch_musa: + if "_MUSA_VISIBLE_DEVICES" in os.environ: + os.environ["MUSA_VISIBLE_DEVICES"] = os.environ["_MUSA_VISIBLE_DEVICES"] + if torch.musa.is_available(): + device = "musa" if os.path.exists("./gweight.txt"): - with open("./gweight.txt", 'r',encoding="utf-8") as file: + with open("./gweight.txt", 'r', encoding="utf-8") as file: gweight_data = file.read() gpt_path = os.environ.get( - "gpt_path", gweight_data) + "gpt_path", gweight_data) else: gpt_path = os.environ.get( - "gpt_path", "GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt") + "gpt_path", "GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt") if os.path.exists("./sweight.txt"): - with open("./sweight.txt", 'r',encoding="utf-8") as file: + with open("./sweight.txt", 'r', encoding="utf-8") as file: sweight_data = file.read() sovits_path = os.environ.get("sovits_path", sweight_data) else: @@ -29,24 +51,25 @@ # ) # sovits_path = os.environ.get("sovits_path", "pretrained_models/s2G488k.pth") cnhubert_base_path = os.environ.get( - "cnhubert_base_path", "pretrained_models/chinese-hubert-base" + "cnhubert_base_path", "GPT_SoVITS/pretrained_models/chinese-hubert-base" ) bert_path = os.environ.get( - "bert_path", "pretrained_models/chinese-roberta-wwm-ext-large" + "bert_path", "GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large" ) infer_ttswebui = os.environ.get("infer_ttswebui", 9872) infer_ttswebui = int(infer_ttswebui) is_share = os.environ.get("is_share", "False") -is_share=eval(is_share) +is_share = eval(is_share) if "_CUDA_VISIBLE_DEVICES" in os.environ: os.environ["CUDA_VISIBLE_DEVICES"] = os.environ["_CUDA_VISIBLE_DEVICES"] -is_half = eval(os.environ.get("is_half", "True")) +is_half = eval(os.environ.get("is_half", "True")) and torch.cuda.is_available() import gradio as gr from transformers import AutoModelForMaskedLM, AutoTokenizer import numpy as np -import librosa,torch +import librosa from feature_extractor import cnhubert -cnhubert.cnhubert_base_path=cnhubert_base_path + +cnhubert.cnhubert_base_path = cnhubert_base_path from module.models import SynthesizerTrn from AR.models.t2s_lightning_module import Text2SemanticLightningModule @@ -56,16 +79,13 @@ from module.mel_processing import spectrogram_torch from my_utils import load_audio from tools.i18n.i18n import I18nAuto + i18n = I18nAuto() -os.environ['PYTORCH_ENABLE_MPS_FALLBACK'] = '1' # 确保直接启动推理UI时也能够设置。 +# os.environ['PYTORCH_ENABLE_MPS_FALLBACK'] = '1' # 确保直接启动推理UI时也能够设置。 if torch.cuda.is_available(): device = "cuda" -elif torch.backends.mps.is_available(): - device = "mps" -else: - device = "cpu" tokenizer = AutoTokenizer.from_pretrained(bert_path) bert_model = AutoModelForMaskedLM.from_pretrained(bert_path) @@ -74,6 +94,7 @@ else: bert_model = bert_model.to(device) + def get_bert_feature(text, word2ph): with torch.no_grad(): inputs = tokenizer(text, return_tensors="pt") @@ -89,6 +110,7 @@ def get_bert_feature(text, word2ph): 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) @@ -123,10 +145,11 @@ def __delattr__(self, item): else: ssl_model = ssl_model.to(device) + def change_sovits_weights(sovits_path): - global vq_model,hps - dict_s2=torch.load(sovits_path,map_location="cpu") - hps=dict_s2["config"] + 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( @@ -135,7 +158,7 @@ def change_sovits_weights(sovits_path): n_speakers=hps.data.n_speakers, **hps.model ) - if("pretrained"not in sovits_path): + if ("pretrained" not in sovits_path): del vq_model.enc_q if is_half == True: vq_model = vq_model.half().to(device) @@ -143,11 +166,15 @@ def change_sovits_weights(sovits_path): vq_model = vq_model.to(device) vq_model.eval() print(vq_model.load_state_dict(dict_s2["weight"], strict=False)) - with open("./sweight.txt","w",encoding="utf-8")as f:f.write(sovits_path) + with open("./sweight.txt", "w", encoding="utf-8") as f: + f.write(sovits_path) + + change_sovits_weights(sovits_path) + def change_gpt_weights(gpt_path): - global hz,max_sec,t2s_model,config + global hz, max_sec, t2s_model, config hz = 50 dict_s1 = torch.load(gpt_path, map_location="cpu") config = dict_s1["config"] @@ -160,9 +187,12 @@ def change_gpt_weights(gpt_path): t2s_model.eval() total = sum([param.nelement() for param in t2s_model.parameters()]) print("Number of parameter: %.2fM" % (total / 1e6)) - with open("./gweight.txt","w",encoding="utf-8")as f:f.write(gpt_path) + with open("./gweight.txt", "w", encoding="utf-8") as f: f.write(gpt_path) + + change_gpt_weights(gpt_path) + def get_spepc(hps, filename): audio = load_audio(filename, int(hps.data.sampling_rate)) audio = torch.FloatTensor(audio) @@ -179,43 +209,26 @@ def get_spepc(hps, filename): return spec -dict_language={ - i18n("中文"):"zh", - i18n("英文"):"en", - i18n("日文"):"ja" +dict_language = { + i18n("中文"): "all_zh",#全部按中文识别 + i18n("英文"): "en",#全部按英文识别#######不变 + i18n("日文"): "all_ja",#全部按日文识别 + i18n("中英混合"): "zh",#按中英混合识别####不变 + i18n("日英混合"): "ja",#按日英混合识别####不变 + i18n("多语种混合"): "auto",#多语种启动切分识别语种 } -def splite_en_inf(sentence, language): - pattern = re.compile(r'[a-zA-Z. ]+') - textlist = [] - langlist = [] - pos = 0 - for match in pattern.finditer(sentence): - start, end = match.span() - if start > pos: - textlist.append(sentence[pos:start]) - langlist.append(language) - textlist.append(sentence[start:end]) - langlist.append("en") - pos = end - if pos < len(sentence): - textlist.append(sentence[pos:]) - langlist.append(language) - - return textlist, langlist - - 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 - +dtype=torch.float16 if is_half == True else torch.float32 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) + bert = get_bert_feature(norm_text, word2ph).to(device)#.to(dtype) else: bert = torch.zeros( (1024, len(phones)), @@ -225,54 +238,112 @@ def get_bert_inf(phones, word2ph, norm_text, language): return bert -def nonen_clean_text_inf(text, language): - textlist, langlist = splite_en_inf(text, language) - phones_list = [] - word2ph_list = [] - norm_text_list = [] - for i in range(len(textlist)): - lang = langlist[i] - phones, word2ph, norm_text = clean_text_inf(textlist[i], lang) - phones_list.append(phones) - if lang == "en" or "ja": - pass - else: - word2ph_list.append(word2ph) - norm_text_list.append(norm_text) - print(word2ph_list) - phones = sum(phones_list, []) - word2ph = sum(word2ph_list, []) - norm_text = ' '.join(norm_text_list) +splits = {"，", "。", "？", "！", ",", ".", "?", "!", "~", ":", "：", "—", "…", } - return phones, word2ph, norm_text +def get_first(text): + pattern = "[" + "".join(re.escape(sep) for sep in splits) + "]" + text = re.split(pattern, text)[0].strip() + return text -def nonen_get_bert_inf(text, language): - textlist, langlist = splite_en_inf(text, language) - print(textlist) - print(langlist) - bert_list = [] - for i in range(len(textlist)): - text = textlist[i] - lang = langlist[i] - phones, word2ph, norm_text = clean_text_inf(text, lang) - bert = get_bert_inf(phones, word2ph, norm_text, lang) - bert_list.append(bert) - bert = torch.cat(bert_list, dim=1) - return bert +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 == True else torch.float32, + ).to(device) + elif language in {"zh", "ja","auto"}: + textlist=[] + langlist=[] + LangSegment.setfilters(["zh","ja","en","ko"]) + if language == "auto": + for tmp in LangSegment.getTexts(text): + if tmp["lang"] == "ko": + langlist.append("zh") + textlist.append(tmp["text"]) + else: + 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 -#i18n("不切"),i18n("凑五句一切"),i18n("凑50字一切"),i18n("按中文句号。切"),i18n("按英文句号.切") -def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language,how_to_cut=i18n("不切")): +def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language, how_to_cut=i18n("不切"), 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 t0 = ttime() - prompt_text = prompt_text.strip("\n") - prompt_language, text = prompt_language, text.strip("\n") + prompt_language = dict_language[prompt_language] + text_language = dict_language[text_language] + if not ref_free: + prompt_text = prompt_text.strip("\n") + if (prompt_text[-1] not in splits): prompt_text += "。" if prompt_language != "en" else "." + print(i18n("实际输入的参考文本:"), prompt_text) + text = text.strip("\n") + if (text[0] not in splits and len(get_first(text)) < 4): text = "。" + text if text_language != "en" else "." + text + + print(i18n("实际输入的目标文本:"), text) zero_wav = np.zeros( int(hps.data.sampling_rate * 0.3), dtype=np.float16 if is_half == True 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(i18n("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 == True: @@ -281,52 +352,51 @@ def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language, else: wav16k = wav16k.to(device) zero_wav_torch = zero_wav_torch.to(device) - wav16k=torch.cat([wav16k,zero_wav_torch]) + 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] t1 = ttime() - prompt_language = dict_language[prompt_language] - text_language = dict_language[text_language] - if prompt_language == "en": - phones1, word2ph1, norm_text1 = clean_text_inf(prompt_text, prompt_language) - else: - phones1, word2ph1, norm_text1 = nonen_clean_text_inf(prompt_text, prompt_language) - if(how_to_cut==i18n("凑五句一切")):text=cut1(text) - elif(how_to_cut==i18n("凑50字一切")):text=cut2(text) - elif(how_to_cut==i18n("按中文句号。切")):text=cut3(text) - elif(how_to_cut==i18n("按英文句号.切")):text=cut4(text) - text = text.replace("\n\n","\n").replace("\n\n","\n").replace("\n\n","\n") - if(text[-1]not in splits):text+="。"if text_language!="en"else "." - texts=text.split("\n") + if (how_to_cut == i18n("凑四句一切")): + text = cut1(text) + elif (how_to_cut == i18n("凑50字一切")): + text = cut2(text) + elif (how_to_cut == i18n("按中文句号。切")): + text = cut3(text) + elif (how_to_cut == i18n("按英文句号.切")): + text = cut4(text) + elif (how_to_cut == i18n("按标点符号切")): + text = cut5(text) + while "\n\n" in text: + text = text.replace("\n\n", "\n") + print(i18n("实际输入的目标文本(切句后):"), text) + texts = text.split("\n") + texts = merge_short_text_in_array(texts, 5) audio_opt = [] - if prompt_language == "en": - bert1 = get_bert_inf(phones1, word2ph1, norm_text1, prompt_language) - else: - bert1 = nonen_get_bert_inf(prompt_text, prompt_language) - + 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_language == "en": - phones2, word2ph2, norm_text2 = clean_text_inf(text, text_language) - else: - phones2, word2ph2, norm_text2 = nonen_clean_text_inf(text, text_language) - - if text_language == "en": - bert2 = get_bert_inf(phones2, word2ph2, norm_text2, text_language) + if (text[-1] not in splits): text += "。" if text_language != "en" else "." + print(i18n("实际输入的目标文本(每句):"), text) + phones2,bert2,norm_text2=get_phones_and_bert(text, text_language) + print(i18n("前端处理后的文本(每句):"), norm_text2) + if not ref_free: + bert = torch.cat([bert1, bert2], 1) + all_phoneme_ids = torch.LongTensor(phones1+phones2).to(device).unsqueeze(0) else: - bert2 = nonen_get_bert_inf(text, text_language) + bert = bert2 + all_phoneme_ids = torch.LongTensor(phones2).to(device).unsqueeze(0) - 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) @@ -336,10 +406,12 @@ def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language, pred_semantic, idx = t2s_model.model.infer_panel( all_phoneme_ids, all_phoneme_len, - prompt, + None if ref_free else prompt, bert, # prompt_phone_len=ph_offset, - top_k=config["inference"]["top_k"], + top_k=top_k, + top_p=top_p, + temperature=temperature, early_stop_num=hz * max_sec, ) t3 = ttime() @@ -357,10 +429,12 @@ def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language, vq_model.decode( pred_semantic, torch.LongTensor(phones2).to(device).unsqueeze(0), refer ) - .detach() - .cpu() - .numpy()[0, 0] + .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) t4 = ttime() @@ -370,23 +444,6 @@ def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language, ) -splits = { - "，", - "。", - "？", - "！", - ",", - ".", - "?", - "!", - "~", - ":", - "：", - "—", - "…", -} # 不考虑省略号 - - def split(todo_text): todo_text = todo_text.replace("……", "。").replace("——", "，") if todo_text[-1] not in splits: @@ -409,12 +466,12 @@ def split(todo_text): def cut1(inp): inp = inp.strip("\n") inps = split(inp) - split_idx = list(range(0, len(inps), 5)) + 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]])) + opts.append("".join(inps[split_idx[idx]: split_idx[idx + 1]])) else: opts = [inp] return "\n".join(opts) @@ -424,7 +481,7 @@ def cut2(inp): inp = inp.strip("\n") inps = split(inp) if len(inps) < 2: - return [inp] + return inp opts = [] summ = 0 tmp_str = "" @@ -437,7 +494,8 @@ def cut2(inp): tmp_str = "" if tmp_str != "": opts.append(tmp_str) - if len(opts[-1]) < 50: ##如果最后一个太短了，和前一个合一起 + # 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) @@ -445,10 +503,28 @@ def cut2(inp): def cut3(inp): inp = inp.strip("\n") - return "\n".join(["%s。" % item for item in inp.strip("。").split("。")]) + 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(".")]) + 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 + def custom_sort_key(s): # 使用正则表达式提取字符串中的数字部分和非数字部分 @@ -457,25 +533,31 @@ def custom_sort_key(s): parts = [int(part) if part.isdigit() else part for part in parts] return parts + def change_choices(): SoVITS_names, GPT_names = get_weights_names() - return {"choices": sorted(SoVITS_names,key=custom_sort_key), "__type__": "update"}, {"choices": sorted(GPT_names,key=custom_sort_key), "__type__": "update"} - -pretrained_sovits_name="GPT_SoVITS/pretrained_models/s2G488k.pth" -pretrained_gpt_name="GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt" -SoVITS_weight_root="SoVITS_weights" -GPT_weight_root="GPT_weights" -os.makedirs(SoVITS_weight_root,exist_ok=True) -os.makedirs(GPT_weight_root,exist_ok=True) + return {"choices": sorted(SoVITS_names, key=custom_sort_key), "__type__": "update"}, {"choices": sorted(GPT_names, key=custom_sort_key), "__type__": "update"} + + +pretrained_sovits_name = "GPT_SoVITS/pretrained_models/s2G488k.pth" +pretrained_gpt_name = "GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt" +SoVITS_weight_root = "SoVITS_weights" +GPT_weight_root = "GPT_weights" +os.makedirs(SoVITS_weight_root, exist_ok=True) +os.makedirs(GPT_weight_root, exist_ok=True) + + def get_weights_names(): SoVITS_names = [pretrained_sovits_name] for name in os.listdir(SoVITS_weight_root): - if name.endswith(".pth"):SoVITS_names.append("%s/%s"%(SoVITS_weight_root,name)) + if name.endswith(".pth"): SoVITS_names.append("%s/%s" % (SoVITS_weight_root, name)) GPT_names = [pretrained_gpt_name] for name in os.listdir(GPT_weight_root): - if name.endswith(".ckpt"): GPT_names.append("%s/%s"%(GPT_weight_root,name)) - return SoVITS_names,GPT_names -SoVITS_names,GPT_names = get_weights_names() + if name.endswith(".ckpt"): GPT_names.append("%s/%s" % (GPT_weight_root, name)) + return SoVITS_names, GPT_names + + +SoVITS_names, GPT_names = get_weights_names() with gr.Blocks(title="GPT-SoVITS WebUI") as app: gr.Markdown( @@ -484,53 +566,63 @@ def get_weights_names(): with gr.Group(): gr.Markdown(value=i18n("模型切换")) with gr.Row(): - GPT_dropdown = gr.Dropdown(label=i18n("GPT模型列表"), choices=sorted(GPT_names, key=custom_sort_key), value=gpt_path,interactive=True) - SoVITS_dropdown = gr.Dropdown(label=i18n("SoVITS模型列表"), choices=sorted(SoVITS_names, key=custom_sort_key), value=sovits_path,interactive=True) + GPT_dropdown = gr.Dropdown(label=i18n("GPT模型列表"), choices=sorted(GPT_names, key=custom_sort_key), value=gpt_path, interactive=True) + SoVITS_dropdown = gr.Dropdown(label=i18n("SoVITS模型列表"), choices=sorted(SoVITS_names, key=custom_sort_key), value=sovits_path, interactive=True) refresh_button = gr.Button(i18n("刷新模型路径"), variant="primary") refresh_button.click(fn=change_choices, inputs=[], outputs=[SoVITS_dropdown, GPT_dropdown]) - SoVITS_dropdown.change(change_sovits_weights,[SoVITS_dropdown],[]) - GPT_dropdown.change(change_gpt_weights,[GPT_dropdown],[]) + SoVITS_dropdown.change(change_sovits_weights, [SoVITS_dropdown], []) + GPT_dropdown.change(change_gpt_weights, [GPT_dropdown], []) gr.Markdown(value=i18n("*请上传并填写参考信息")) with gr.Row(): - inp_ref = gr.Audio(label=i18n("请上传参考音频"), type="filepath") - prompt_text = gr.Textbox(label=i18n("参考音频的文本"), value="") + inp_ref = gr.Audio(label=i18n("请上传3~10秒内参考音频，超过会报错！"), type="filepath") + with gr.Column(): + ref_text_free = gr.Checkbox(label=i18n("开启无参考文本模式。不填参考文本亦相当于开启。"), value=False, interactive=True, show_label=True) + gr.Markdown(i18n("使用无参考文本模式时建议使用微调的GPT，听不清参考音频说的啥(不晓得写啥)可以开，开启后无视填写的参考文本。")) + prompt_text = gr.Textbox(label=i18n("参考音频的文本"), value="") prompt_language = gr.Dropdown( - label=i18n("参考音频的语种"),choices=[i18n("中文"),i18n("英文"),i18n("日文")],value=i18n("中文") + label=i18n("参考音频的语种"), choices=[i18n("中文"), i18n("英文"), i18n("日文"), i18n("中英混合"), i18n("日英混合"), i18n("多语种混合")], value=i18n("中文") ) - gr.Markdown(value=i18n("*请填写需要合成的目标文本。中英混合选中文，日英混合选日文，中日混合暂不支持，非目标语言文本自动遗弃。")) + gr.Markdown(value=i18n("*请填写需要合成的目标文本和语种模式")) with gr.Row(): text = gr.Textbox(label=i18n("需要合成的文本"), value="") text_language = gr.Dropdown( - label=i18n("需要合成的语种"),choices=[i18n("中文"),i18n("英文"),i18n("日文")],value=i18n("中文") + label=i18n("需要合成的语种"), choices=[i18n("中文"), i18n("英文"), i18n("日文"), i18n("中英混合"), i18n("日英混合"), i18n("多语种混合")], value=i18n("中文") ) how_to_cut = gr.Radio( label=i18n("怎么切"), - choices=[i18n("不切"),i18n("凑五句一切"),i18n("凑50字一切"),i18n("按中文句号。切"),i18n("按英文句号.切"),], - value=i18n("凑50字一切"), + choices=[i18n("不切"), i18n("凑四句一切"), i18n("凑50字一切"), i18n("按中文句号。切"), i18n("按英文句号.切"), i18n("按标点符号切"), ], + value=i18n("凑四句一切"), interactive=True, ) + with gr.Row(): + gr.Markdown(value=i18n("gpt采样参数(无参考文本时不要太低)：")) + top_k = gr.Slider(minimum=1,maximum=100,step=1,label=i18n("top_k"),value=5,interactive=True) + top_p = gr.Slider(minimum=0,maximum=1,step=0.05,label=i18n("top_p"),value=1,interactive=True) + temperature = gr.Slider(minimum=0,maximum=1,step=0.05,label=i18n("temperature"),value=1,interactive=True) inference_button = gr.Button(i18n("合成语音"), variant="primary") output = gr.Audio(label=i18n("输出的语音")) inference_button.click( get_tts_wav, - [inp_ref, prompt_text, prompt_language, text, text_language,how_to_cut], + [inp_ref, prompt_text, prompt_language, text, text_language, how_to_cut, top_k, top_p, temperature, ref_text_free], [output], ) gr.Markdown(value=i18n("文本切分工具。太长的文本合成出来效果不一定好，所以太长建议先切。合成会根据文本的换行分开合成再拼起来。")) with gr.Row(): - text_inp = gr.Textbox(label=i18n("需要合成的切分前文本"),value="") - button1 = gr.Button(i18n("凑五句一切"), variant="primary") + text_inp = gr.Textbox(label=i18n("需要合成的切分前文本"), value="") + button1 = gr.Button(i18n("凑四句一切"), variant="primary") button2 = gr.Button(i18n("凑50字一切"), variant="primary") button3 = gr.Button(i18n("按中文句号。切"), variant="primary") button4 = gr.Button(i18n("按英文句号.切"), variant="primary") + button5 = gr.Button(i18n("按标点符号切"), variant="primary") text_opt = gr.Textbox(label=i18n("切分后文本"), value="") button1.click(cut1, [text_inp], [text_opt]) button2.click(cut2, [text_inp], [text_opt]) button3.click(cut3, [text_inp], [text_opt]) button4.click(cut4, [text_inp], [text_opt]) - gr.Markdown(value=i18n("后续将支持混合语种编码文本输入。")) + button5.click(cut5, [text_inp], [text_opt]) + gr.Markdown(value=i18n("后续将支持转音素、手工修改音素、语音合成分步执行。")) app.queue(concurrency_count=511, max_size=1022).launch( server_name="0.0.0.0", From f7d3d32cf0967dc9524fb61e2f8725670743112c Mon Sep 17 00:00:00 2001 From: KakaruHayate Date: Sat, 30 Mar 2024 22:14:45 +0800 Subject: [PATCH 05/10] =?UTF-8?q?=E5=9C=A8=E5=85=B6=E4=BB=96=E6=8E=A8?= =?UTF-8?q?=E7=90=86=E9=83=A8=E5=88=86=E5=8A=A0=E8=BD=BDtorch=5Fmusa?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- GPT_SoVITS/AR/models/t2s_lightning_module.py | 4 + GPT_SoVITS/AR/models/t2s_model.py | 4 + GPT_SoVITS/AR/models/utils.py | 4 + GPT_SoVITS/AR/modules/activation.py | 4 + GPT_SoVITS/AR/modules/embedding.py | 4 + .../AR/modules/patched_mha_with_cache.py | 4 + GPT_SoVITS/AR/modules/scaling.py | 4 + GPT_SoVITS/AR/modules/transformer.py | 4 + GPT_SoVITS/feature_extractor/cnhubert.py | 4 + GPT_SoVITS/inference_webui.py | 1266 ++++++++--------- GPT_SoVITS/module/attentions.py | 4 + GPT_SoVITS/module/commons.py | 4 + GPT_SoVITS/module/core_vq.py | 4 + GPT_SoVITS/module/mel_processing.py | 4 + GPT_SoVITS/module/models.py | 4 + GPT_SoVITS/module/modules.py | 4 + GPT_SoVITS/module/mrte_model.py | 4 + GPT_SoVITS/module/quantize.py | 4 + GPT_SoVITS/module/transforms.py | 4 + 19 files changed, 705 insertions(+), 633 deletions(-) diff --git a/GPT_SoVITS/AR/models/t2s_lightning_module.py b/GPT_SoVITS/AR/models/t2s_lightning_module.py index 2dd3f392..adbb7206 100644 --- a/GPT_SoVITS/AR/models/t2s_lightning_module.py +++ b/GPT_SoVITS/AR/models/t2s_lightning_module.py @@ -7,6 +7,10 @@ from typing import Dict import torch +try: + import torch_musa +except ImportError: + pass from pytorch_lightning import LightningModule from AR.models.t2s_model import Text2SemanticDecoder from AR.modules.lr_schedulers import WarmupCosineLRSchedule diff --git a/GPT_SoVITS/AR/models/t2s_model.py b/GPT_SoVITS/AR/models/t2s_model.py index c8ad3d82..5805a6ce 100644 --- a/GPT_SoVITS/AR/models/t2s_model.py +++ b/GPT_SoVITS/AR/models/t2s_model.py @@ -1,6 +1,10 @@ # modified from https://github.com/yangdongchao/SoundStorm/blob/master/soundstorm/s1/AR/models/t2s_model.py # reference: https://github.com/lifeiteng/vall-e import torch +try: + import torch_musa +except ImportError: + pass from tqdm import tqdm from AR.models.utils import make_pad_mask diff --git a/GPT_SoVITS/AR/models/utils.py b/GPT_SoVITS/AR/models/utils.py index 9678c7e1..706338ea 100644 --- a/GPT_SoVITS/AR/models/utils.py +++ b/GPT_SoVITS/AR/models/utils.py @@ -1,6 +1,10 @@ # modified from https://github.com/yangdongchao/SoundStorm/blob/master/soundstorm/s1/AR/models/utils.py # reference: https://github.com/lifeiteng/vall-e import torch +try: + import torch_musa +except ImportError: + pass import torch.nn.functional as F from typing import Tuple diff --git a/GPT_SoVITS/AR/modules/activation.py b/GPT_SoVITS/AR/modules/activation.py index 5ca888b5..89b4cddf 100644 --- a/GPT_SoVITS/AR/modules/activation.py +++ b/GPT_SoVITS/AR/modules/activation.py @@ -2,6 +2,10 @@ from typing import Optional from typing import Tuple import torch +try: + import torch_musa +except ImportError: + pass from torch import Tensor from torch.nn import Linear from torch.nn import Module diff --git a/GPT_SoVITS/AR/modules/embedding.py b/GPT_SoVITS/AR/modules/embedding.py index 3a382f93..f1736bd3 100644 --- a/GPT_SoVITS/AR/modules/embedding.py +++ b/GPT_SoVITS/AR/modules/embedding.py @@ -2,6 +2,10 @@ import math import torch +try: + import torch_musa +except ImportError: + pass from torch import nn diff --git a/GPT_SoVITS/AR/modules/patched_mha_with_cache.py b/GPT_SoVITS/AR/modules/patched_mha_with_cache.py index 7be241da..9c4729b4 100644 --- a/GPT_SoVITS/AR/modules/patched_mha_with_cache.py +++ b/GPT_SoVITS/AR/modules/patched_mha_with_cache.py @@ -7,6 +7,10 @@ ) from torch.nn import functional as F import torch +try: + import torch_musa +except ImportError: + pass # Tensor = torch.Tensor # from typing import Callable, List, Optional, Tuple, Union diff --git a/GPT_SoVITS/AR/modules/scaling.py b/GPT_SoVITS/AR/modules/scaling.py index 9256a8cb..20679d63 100644 --- a/GPT_SoVITS/AR/modules/scaling.py +++ b/GPT_SoVITS/AR/modules/scaling.py @@ -21,6 +21,10 @@ from typing import Union import torch +try: + import torch_musa +except ImportError: + pass import torch.nn as nn from torch import Tensor diff --git a/GPT_SoVITS/AR/modules/transformer.py b/GPT_SoVITS/AR/modules/transformer.py index 7921f48e..4d3d286c 100644 --- a/GPT_SoVITS/AR/modules/transformer.py +++ b/GPT_SoVITS/AR/modules/transformer.py @@ -10,6 +10,10 @@ from typing import Union import torch +try: + import torch_musa +except ImportError: + pass from AR.modules.activation import MultiheadAttention from AR.modules.scaling import BalancedDoubleSwish from torch import nn diff --git a/GPT_SoVITS/feature_extractor/cnhubert.py b/GPT_SoVITS/feature_extractor/cnhubert.py index dc155bdd..a23293f1 100644 --- a/GPT_SoVITS/feature_extractor/cnhubert.py +++ b/GPT_SoVITS/feature_extractor/cnhubert.py @@ -2,6 +2,10 @@ import librosa import torch +try: + import torch_musa +except ImportError: + pass import torch.nn.functional as F import soundfile as sf import logging diff --git a/GPT_SoVITS/inference_webui.py b/GPT_SoVITS/inference_webui.py index 237d1b50..48541c17 100644 --- a/GPT_SoVITS/inference_webui.py +++ b/GPT_SoVITS/inference_webui.py @@ -1,633 +1,633 @@ -''' -按中英混合识别 -按日英混合识别 -多语种启动切分识别语种 -全部按中文识别 -全部按英文识别 -全部按日文识别 -''' -import os, re, logging -import LangSegment -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) -import pdb -import torch - -device = "cpu" - -try: - import torch_musa - use_torch_musa = True -except ImportError: - use_torch_musa = False -if use_torch_musa: - if "_MUSA_VISIBLE_DEVICES" in os.environ: - os.environ["MUSA_VISIBLE_DEVICES"] = os.environ["_MUSA_VISIBLE_DEVICES"] - if torch.musa.is_available(): - device = "musa" - -if os.path.exists("./gweight.txt"): - with open("./gweight.txt", 'r', encoding="utf-8") as file: - gweight_data = file.read() - gpt_path = os.environ.get( - "gpt_path", gweight_data) -else: - gpt_path = os.environ.get( - "gpt_path", "GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt") - -if os.path.exists("./sweight.txt"): - with open("./sweight.txt", 'r', encoding="utf-8") as file: - sweight_data = file.read() - sovits_path = os.environ.get("sovits_path", sweight_data) -else: - sovits_path = os.environ.get("sovits_path", "GPT_SoVITS/pretrained_models/s2G488k.pth") -# gpt_path = os.environ.get( -# "gpt_path", "pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt" -# ) -# sovits_path = os.environ.get("sovits_path", "pretrained_models/s2G488k.pth") -cnhubert_base_path = os.environ.get( - "cnhubert_base_path", "GPT_SoVITS/pretrained_models/chinese-hubert-base" -) -bert_path = os.environ.get( - "bert_path", "GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large" -) -infer_ttswebui = os.environ.get("infer_ttswebui", 9872) -infer_ttswebui = int(infer_ttswebui) -is_share = os.environ.get("is_share", "False") -is_share = eval(is_share) -if "_CUDA_VISIBLE_DEVICES" in os.environ: - os.environ["CUDA_VISIBLE_DEVICES"] = os.environ["_CUDA_VISIBLE_DEVICES"] -is_half = eval(os.environ.get("is_half", "True")) and torch.cuda.is_available() -import gradio as gr -from transformers import AutoModelForMaskedLM, AutoTokenizer -import numpy as np -import librosa -from feature_extractor import cnhubert - -cnhubert.cnhubert_base_path = cnhubert_base_path - -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 time import time as ttime -from module.mel_processing import spectrogram_torch -from my_utils import load_audio -from tools.i18n.i18n import I18nAuto - -i18n = I18nAuto() - -# os.environ['PYTORCH_ENABLE_MPS_FALLBACK'] = '1' # 确保直接启动推理UI时也能够设置。 - -if torch.cuda.is_available(): - device = "cuda" - -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) - - -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") - - -ssl_model = cnhubert.get_model() -if is_half == True: - ssl_model = ssl_model.half().to(device) -else: - ssl_model = ssl_model.to(device) - - -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 == True: - vq_model = vq_model.half().to(device) - else: - vq_model = vq_model.to(device) - vq_model.eval() - print(vq_model.load_state_dict(dict_s2["weight"], strict=False)) - with open("./sweight.txt", "w", encoding="utf-8") as f: - f.write(sovits_path) - - -change_sovits_weights(sovits_path) - - -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 == 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)) - with open("./gweight.txt", "w", encoding="utf-8") as f: f.write(gpt_path) - - -change_gpt_weights(gpt_path) - - -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 - - -dict_language = { - i18n("中文"): "all_zh",#全部按中文识别 - i18n("英文"): "en",#全部按英文识别#######不变 - i18n("日文"): "all_ja",#全部按日文识别 - i18n("中英混合"): "zh",#按中英混合识别####不变 - i18n("日英混合"): "ja",#按日英混合识别####不变 - i18n("多语种混合"): "auto",#多语种启动切分识别语种 -} - - -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 - -dtype=torch.float16 if is_half == True else torch.float32 -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 == True else torch.float32, - ).to(device) - - return bert - - -splits = {"，", "。", "？", "！", ",", ".", "?", "!", "~", ":", "：", "—", "…", } - - -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 == True else torch.float32, - ).to(device) - elif language in {"zh", "ja","auto"}: - textlist=[] - langlist=[] - LangSegment.setfilters(["zh","ja","en","ko"]) - if language == "auto": - for tmp in LangSegment.getTexts(text): - if tmp["lang"] == "ko": - langlist.append("zh") - textlist.append(tmp["text"]) - else: - 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=i18n("不切"), 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 - t0 = ttime() - prompt_language = dict_language[prompt_language] - text_language = dict_language[text_language] - if not ref_free: - prompt_text = prompt_text.strip("\n") - if (prompt_text[-1] not in splits): prompt_text += "。" if prompt_language != "en" else "." - print(i18n("实际输入的参考文本:"), prompt_text) - text = text.strip("\n") - if (text[0] not in splits and len(get_first(text)) < 4): text = "。" + text if text_language != "en" else "." + text - - print(i18n("实际输入的目标文本:"), text) - zero_wav = np.zeros( - int(hps.data.sampling_rate * 0.3), - dtype=np.float16 if is_half == True 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(i18n("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 == True: - 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] - t1 = ttime() - - if (how_to_cut == i18n("凑四句一切")): - text = cut1(text) - elif (how_to_cut == i18n("凑50字一切")): - text = cut2(text) - elif (how_to_cut == i18n("按中文句号。切")): - text = cut3(text) - elif (how_to_cut == i18n("按英文句号.切")): - text = cut4(text) - elif (how_to_cut == i18n("按标点符号切")): - text = cut5(text) - while "\n\n" in text: - text = text.replace("\n\n", "\n") - print(i18n("实际输入的目标文本(切句后):"), text) - 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 "." - print(i18n("实际输入的目标文本(每句):"), text) - phones2,bert2,norm_text2=get_phones_and_bert(text, text_language) - print(i18n("前端处理后的文本(每句):"), norm_text2) - 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) - 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, - 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, - ) - 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 - 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) - t4 = ttime() - print("%.3f\t%.3f\t%.3f\t%.3f" % (t1 - t0, t2 - t1, t3 - t2, t4 - t3)) - 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 - - -def custom_sort_key(s): - # 使用正则表达式提取字符串中的数字部分和非数字部分 - parts = re.split('(\d+)', s) - # 将数字部分转换为整数，非数字部分保持不变 - parts = [int(part) if part.isdigit() else part for part in parts] - return parts - - -def change_choices(): - SoVITS_names, GPT_names = get_weights_names() - return {"choices": sorted(SoVITS_names, key=custom_sort_key), "__type__": "update"}, {"choices": sorted(GPT_names, key=custom_sort_key), "__type__": "update"} - - -pretrained_sovits_name = "GPT_SoVITS/pretrained_models/s2G488k.pth" -pretrained_gpt_name = "GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt" -SoVITS_weight_root = "SoVITS_weights" -GPT_weight_root = "GPT_weights" -os.makedirs(SoVITS_weight_root, exist_ok=True) -os.makedirs(GPT_weight_root, exist_ok=True) - - -def get_weights_names(): - SoVITS_names = [pretrained_sovits_name] - for name in os.listdir(SoVITS_weight_root): - if name.endswith(".pth"): SoVITS_names.append("%s/%s" % (SoVITS_weight_root, name)) - GPT_names = [pretrained_gpt_name] - for name in os.listdir(GPT_weight_root): - if name.endswith(".ckpt"): GPT_names.append("%s/%s" % (GPT_weight_root, name)) - return SoVITS_names, GPT_names - - -SoVITS_names, GPT_names = get_weights_names() - -with gr.Blocks(title="GPT-SoVITS WebUI") as app: - gr.Markdown( - value=i18n("本软件以MIT协议开源, 作者不对软件具备任何控制力, 使用软件者、传播软件导出的声音者自负全责.
如不认可该条款, 则不能使用或引用软件包内任何代码和文件. 详见根目录LICENSE.") - ) - with gr.Group(): - gr.Markdown(value=i18n("模型切换")) - with gr.Row(): - GPT_dropdown = gr.Dropdown(label=i18n("GPT模型列表"), choices=sorted(GPT_names, key=custom_sort_key), value=gpt_path, interactive=True) - SoVITS_dropdown = gr.Dropdown(label=i18n("SoVITS模型列表"), choices=sorted(SoVITS_names, key=custom_sort_key), value=sovits_path, interactive=True) - refresh_button = gr.Button(i18n("刷新模型路径"), variant="primary") - refresh_button.click(fn=change_choices, inputs=[], outputs=[SoVITS_dropdown, GPT_dropdown]) - SoVITS_dropdown.change(change_sovits_weights, [SoVITS_dropdown], []) - GPT_dropdown.change(change_gpt_weights, [GPT_dropdown], []) - gr.Markdown(value=i18n("*请上传并填写参考信息")) - with gr.Row(): - inp_ref = gr.Audio(label=i18n("请上传3~10秒内参考音频，超过会报错！"), type="filepath") - with gr.Column(): - ref_text_free = gr.Checkbox(label=i18n("开启无参考文本模式。不填参考文本亦相当于开启。"), value=False, interactive=True, show_label=True) - gr.Markdown(i18n("使用无参考文本模式时建议使用微调的GPT，听不清参考音频说的啥(不晓得写啥)可以开，开启后无视填写的参考文本。")) - prompt_text = gr.Textbox(label=i18n("参考音频的文本"), value="") - prompt_language = gr.Dropdown( - label=i18n("参考音频的语种"), choices=[i18n("中文"), i18n("英文"), i18n("日文"), i18n("中英混合"), i18n("日英混合"), i18n("多语种混合")], value=i18n("中文") - ) - gr.Markdown(value=i18n("*请填写需要合成的目标文本和语种模式")) - with gr.Row(): - text = gr.Textbox(label=i18n("需要合成的文本"), value="") - text_language = gr.Dropdown( - label=i18n("需要合成的语种"), choices=[i18n("中文"), i18n("英文"), i18n("日文"), i18n("中英混合"), i18n("日英混合"), i18n("多语种混合")], value=i18n("中文") - ) - how_to_cut = gr.Radio( - label=i18n("怎么切"), - choices=[i18n("不切"), i18n("凑四句一切"), i18n("凑50字一切"), i18n("按中文句号。切"), i18n("按英文句号.切"), i18n("按标点符号切"), ], - value=i18n("凑四句一切"), - interactive=True, - ) - with gr.Row(): - gr.Markdown(value=i18n("gpt采样参数(无参考文本时不要太低)：")) - top_k = gr.Slider(minimum=1,maximum=100,step=1,label=i18n("top_k"),value=5,interactive=True) - top_p = gr.Slider(minimum=0,maximum=1,step=0.05,label=i18n("top_p"),value=1,interactive=True) - temperature = gr.Slider(minimum=0,maximum=1,step=0.05,label=i18n("temperature"),value=1,interactive=True) - inference_button = gr.Button(i18n("合成语音"), variant="primary") - output = gr.Audio(label=i18n("输出的语音")) - - inference_button.click( - get_tts_wav, - [inp_ref, prompt_text, prompt_language, text, text_language, how_to_cut, top_k, top_p, temperature, ref_text_free], - [output], - ) - - gr.Markdown(value=i18n("文本切分工具。太长的文本合成出来效果不一定好，所以太长建议先切。合成会根据文本的换行分开合成再拼起来。")) - with gr.Row(): - text_inp = gr.Textbox(label=i18n("需要合成的切分前文本"), value="") - button1 = gr.Button(i18n("凑四句一切"), variant="primary") - button2 = gr.Button(i18n("凑50字一切"), variant="primary") - button3 = gr.Button(i18n("按中文句号。切"), variant="primary") - button4 = gr.Button(i18n("按英文句号.切"), variant="primary") - button5 = gr.Button(i18n("按标点符号切"), variant="primary") - text_opt = gr.Textbox(label=i18n("切分后文本"), value="") - button1.click(cut1, [text_inp], [text_opt]) - button2.click(cut2, [text_inp], [text_opt]) - button3.click(cut3, [text_inp], [text_opt]) - button4.click(cut4, [text_inp], [text_opt]) - button5.click(cut5, [text_inp], [text_opt]) - gr.Markdown(value=i18n("后续将支持转音素、手工修改音素、语音合成分步执行。")) - -app.queue(concurrency_count=511, max_size=1022).launch( - server_name="0.0.0.0", - inbrowser=True, - share=is_share, - server_port=infer_ttswebui, - quiet=True, -) \ No newline at end of file +''' +按中英混合识别 +按日英混合识别 +多语种启动切分识别语种 +全部按中文识别 +全部按英文识别 +全部按日文识别 +''' +import os, re, logging +import LangSegment +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) +import pdb +import torch + +device = "cpu" + +try: + import torch_musa + use_torch_musa = True +except ImportError: + use_torch_musa = False +if use_torch_musa: + if "_MUSA_VISIBLE_DEVICES" in os.environ: + os.environ["MUSA_VISIBLE_DEVICES"] = os.environ["_MUSA_VISIBLE_DEVICES"] + if torch.musa.is_available(): + device = "musa" + +if os.path.exists("./gweight.txt"): + with open("./gweight.txt", 'r', encoding="utf-8") as file: + gweight_data = file.read() + gpt_path = os.environ.get( + "gpt_path", gweight_data) +else: + gpt_path = os.environ.get( + "gpt_path", "GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt") + +if os.path.exists("./sweight.txt"): + with open("./sweight.txt", 'r', encoding="utf-8") as file: + sweight_data = file.read() + sovits_path = os.environ.get("sovits_path", sweight_data) +else: + sovits_path = os.environ.get("sovits_path", "GPT_SoVITS/pretrained_models/s2G488k.pth") +# gpt_path = os.environ.get( +# "gpt_path", "pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt" +# ) +# sovits_path = os.environ.get("sovits_path", "pretrained_models/s2G488k.pth") +cnhubert_base_path = os.environ.get( + "cnhubert_base_path", "GPT_SoVITS/pretrained_models/chinese-hubert-base" +) +bert_path = os.environ.get( + "bert_path", "GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large" +) +infer_ttswebui = os.environ.get("infer_ttswebui", 9872) +infer_ttswebui = int(infer_ttswebui) +is_share = os.environ.get("is_share", "False") +is_share = eval(is_share) +if "_CUDA_VISIBLE_DEVICES" in os.environ: + os.environ["CUDA_VISIBLE_DEVICES"] = os.environ["_CUDA_VISIBLE_DEVICES"] +is_half = eval(os.environ.get("is_half", "True")) and torch.cuda.is_available() +import gradio as gr +from transformers import AutoModelForMaskedLM, AutoTokenizer +import numpy as np +import librosa +from feature_extractor import cnhubert + +cnhubert.cnhubert_base_path = cnhubert_base_path + +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 time import time as ttime +from module.mel_processing import spectrogram_torch +from my_utils import load_audio +from tools.i18n.i18n import I18nAuto + +i18n = I18nAuto() + +# os.environ['PYTORCH_ENABLE_MPS_FALLBACK'] = '1' # 确保直接启动推理UI时也能够设置。 + +if torch.cuda.is_available(): + device = "cuda" + +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) + + +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") + + +ssl_model = cnhubert.get_model() +if is_half == True: + ssl_model = ssl_model.half().to(device) +else: + ssl_model = ssl_model.to(device) + + +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 == True: + vq_model = vq_model.half().to(device) + else: + vq_model = vq_model.to(device) + vq_model.eval() + print(vq_model.load_state_dict(dict_s2["weight"], strict=False)) + with open("./sweight.txt", "w", encoding="utf-8") as f: + f.write(sovits_path) + + +change_sovits_weights(sovits_path) + + +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 == 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)) + with open("./gweight.txt", "w", encoding="utf-8") as f: f.write(gpt_path) + + +change_gpt_weights(gpt_path) + + +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 + + +dict_language = { + i18n("中文"): "all_zh",#全部按中文识别 + i18n("英文"): "en",#全部按英文识别#######不变 + i18n("日文"): "all_ja",#全部按日文识别 + i18n("中英混合"): "zh",#按中英混合识别####不变 + i18n("日英混合"): "ja",#按日英混合识别####不变 + i18n("多语种混合"): "auto",#多语种启动切分识别语种 +} + + +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 + +dtype=torch.float16 if is_half == True else torch.float32 +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 == True else torch.float32, + ).to(device) + + return bert + + +splits = {"，", "。", "？", "！", ",", ".", "?", "!", "~", ":", "：", "—", "…", } + + +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 == True else torch.float32, + ).to(device) + elif language in {"zh", "ja","auto"}: + textlist=[] + langlist=[] + LangSegment.setfilters(["zh","ja","en","ko"]) + if language == "auto": + for tmp in LangSegment.getTexts(text): + if tmp["lang"] == "ko": + langlist.append("zh") + textlist.append(tmp["text"]) + else: + 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=i18n("不切"), 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 + t0 = ttime() + prompt_language = dict_language[prompt_language] + text_language = dict_language[text_language] + if not ref_free: + prompt_text = prompt_text.strip("\n") + if (prompt_text[-1] not in splits): prompt_text += "。" if prompt_language != "en" else "." + print(i18n("实际输入的参考文本:"), prompt_text) + text = text.strip("\n") + if (text[0] not in splits and len(get_first(text)) < 4): text = "。" + text if text_language != "en" else "." + text + + print(i18n("实际输入的目标文本:"), text) + zero_wav = np.zeros( + int(hps.data.sampling_rate * 0.3), + dtype=np.float16 if is_half == True 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(i18n("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 == True: + 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] + t1 = ttime() + + if (how_to_cut == i18n("凑四句一切")): + text = cut1(text) + elif (how_to_cut == i18n("凑50字一切")): + text = cut2(text) + elif (how_to_cut == i18n("按中文句号。切")): + text = cut3(text) + elif (how_to_cut == i18n("按英文句号.切")): + text = cut4(text) + elif (how_to_cut == i18n("按标点符号切")): + text = cut5(text) + while "\n\n" in text: + text = text.replace("\n\n", "\n") + print(i18n("实际输入的目标文本(切句后):"), text) + 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 "." + print(i18n("实际输入的目标文本(每句):"), text) + phones2,bert2,norm_text2=get_phones_and_bert(text, text_language) + print(i18n("前端处理后的文本(每句):"), norm_text2) + 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) + 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, + 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, + ) + 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 + 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) + t4 = ttime() + print("%.3f\t%.3f\t%.3f\t%.3f" % (t1 - t0, t2 - t1, t3 - t2, t4 - t3)) + 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 + + +def custom_sort_key(s): + # 使用正则表达式提取字符串中的数字部分和非数字部分 + parts = re.split('(\d+)', s) + # 将数字部分转换为整数，非数字部分保持不变 + parts = [int(part) if part.isdigit() else part for part in parts] + return parts + + +def change_choices(): + SoVITS_names, GPT_names = get_weights_names() + return {"choices": sorted(SoVITS_names, key=custom_sort_key), "__type__": "update"}, {"choices": sorted(GPT_names, key=custom_sort_key), "__type__": "update"} + + +pretrained_sovits_name = "GPT_SoVITS/pretrained_models/s2G488k.pth" +pretrained_gpt_name = "GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt" +SoVITS_weight_root = "SoVITS_weights" +GPT_weight_root = "GPT_weights" +os.makedirs(SoVITS_weight_root, exist_ok=True) +os.makedirs(GPT_weight_root, exist_ok=True) + + +def get_weights_names(): + SoVITS_names = [pretrained_sovits_name] + for name in os.listdir(SoVITS_weight_root): + if name.endswith(".pth"): SoVITS_names.append("%s/%s" % (SoVITS_weight_root, name)) + GPT_names = [pretrained_gpt_name] + for name in os.listdir(GPT_weight_root): + if name.endswith(".ckpt"): GPT_names.append("%s/%s" % (GPT_weight_root, name)) + return SoVITS_names, GPT_names + + +SoVITS_names, GPT_names = get_weights_names() + +with gr.Blocks(title="GPT-SoVITS WebUI") as app: + gr.Markdown( + value=i18n("本软件以MIT协议开源, 作者不对软件具备任何控制力, 使用软件者、传播软件导出的声音者自负全责.
如不认可该条款, 则不能使用或引用软件包内任何代码和文件. 详见根目录LICENSE.") + ) + with gr.Group(): + gr.Markdown(value=i18n("模型切换")) + with gr.Row(): + GPT_dropdown = gr.Dropdown(label=i18n("GPT模型列表"), choices=sorted(GPT_names, key=custom_sort_key), value=gpt_path, interactive=True) + SoVITS_dropdown = gr.Dropdown(label=i18n("SoVITS模型列表"), choices=sorted(SoVITS_names, key=custom_sort_key), value=sovits_path, interactive=True) + refresh_button = gr.Button(i18n("刷新模型路径"), variant="primary") + refresh_button.click(fn=change_choices, inputs=[], outputs=[SoVITS_dropdown, GPT_dropdown]) + SoVITS_dropdown.change(change_sovits_weights, [SoVITS_dropdown], []) + GPT_dropdown.change(change_gpt_weights, [GPT_dropdown], []) + gr.Markdown(value=i18n("*请上传并填写参考信息")) + with gr.Row(): + inp_ref = gr.Audio(label=i18n("请上传3~10秒内参考音频，超过会报错！"), type="filepath") + with gr.Column(): + ref_text_free = gr.Checkbox(label=i18n("开启无参考文本模式。不填参考文本亦相当于开启。"), value=False, interactive=True, show_label=True) + gr.Markdown(i18n("使用无参考文本模式时建议使用微调的GPT，听不清参考音频说的啥(不晓得写啥)可以开，开启后无视填写的参考文本。")) + prompt_text = gr.Textbox(label=i18n("参考音频的文本"), value="") + prompt_language = gr.Dropdown( + label=i18n("参考音频的语种"), choices=[i18n("中文"), i18n("英文"), i18n("日文"), i18n("中英混合"), i18n("日英混合"), i18n("多语种混合")], value=i18n("中文") + ) + gr.Markdown(value=i18n("*请填写需要合成的目标文本和语种模式")) + with gr.Row(): + text = gr.Textbox(label=i18n("需要合成的文本"), value="") + text_language = gr.Dropdown( + label=i18n("需要合成的语种"), choices=[i18n("中文"), i18n("英文"), i18n("日文"), i18n("中英混合"), i18n("日英混合"), i18n("多语种混合")], value=i18n("中文") + ) + how_to_cut = gr.Radio( + label=i18n("怎么切"), + choices=[i18n("不切"), i18n("凑四句一切"), i18n("凑50字一切"), i18n("按中文句号。切"), i18n("按英文句号.切"), i18n("按标点符号切"), ], + value=i18n("凑四句一切"), + interactive=True, + ) + with gr.Row(): + gr.Markdown(value=i18n("gpt采样参数(无参考文本时不要太低)：")) + top_k = gr.Slider(minimum=1,maximum=100,step=1,label=i18n("top_k"),value=5,interactive=True) + top_p = gr.Slider(minimum=0,maximum=1,step=0.05,label=i18n("top_p"),value=1,interactive=True) + temperature = gr.Slider(minimum=0,maximum=1,step=0.05,label=i18n("temperature"),value=1,interactive=True) + inference_button = gr.Button(i18n("合成语音"), variant="primary") + output = gr.Audio(label=i18n("输出的语音")) + + inference_button.click( + get_tts_wav, + [inp_ref, prompt_text, prompt_language, text, text_language, how_to_cut, top_k, top_p, temperature, ref_text_free], + [output], + ) + + gr.Markdown(value=i18n("文本切分工具。太长的文本合成出来效果不一定好，所以太长建议先切。合成会根据文本的换行分开合成再拼起来。")) + with gr.Row(): + text_inp = gr.Textbox(label=i18n("需要合成的切分前文本"), value="") + button1 = gr.Button(i18n("凑四句一切"), variant="primary") + button2 = gr.Button(i18n("凑50字一切"), variant="primary") + button3 = gr.Button(i18n("按中文句号。切"), variant="primary") + button4 = gr.Button(i18n("按英文句号.切"), variant="primary") + button5 = gr.Button(i18n("按标点符号切"), variant="primary") + text_opt = gr.Textbox(label=i18n("切分后文本"), value="") + button1.click(cut1, [text_inp], [text_opt]) + button2.click(cut2, [text_inp], [text_opt]) + button3.click(cut3, [text_inp], [text_opt]) + button4.click(cut4, [text_inp], [text_opt]) + button5.click(cut5, [text_inp], [text_opt]) + gr.Markdown(value=i18n("后续将支持转音素、手工修改音素、语音合成分步执行。")) + +app.queue(concurrency_count=511, max_size=1022).launch( + server_name="0.0.0.0", + inbrowser=True, + share=is_share, + server_port=infer_ttswebui, + quiet=True, +) diff --git a/GPT_SoVITS/module/attentions.py b/GPT_SoVITS/module/attentions.py index a2e9e515..1f57d890 100644 --- a/GPT_SoVITS/module/attentions.py +++ b/GPT_SoVITS/module/attentions.py @@ -1,5 +1,9 @@ import math import torch +try: + import torch_musa +except ImportError: + pass from torch import nn from torch.nn import functional as F diff --git a/GPT_SoVITS/module/commons.py b/GPT_SoVITS/module/commons.py index e96cf923..5e7703d1 100644 --- a/GPT_SoVITS/module/commons.py +++ b/GPT_SoVITS/module/commons.py @@ -1,5 +1,9 @@ import math import torch +try: + import torch_musa +except ImportError: + pass from torch.nn import functional as F diff --git a/GPT_SoVITS/module/core_vq.py b/GPT_SoVITS/module/core_vq.py index a5e22d66..557ec927 100644 --- a/GPT_SoVITS/module/core_vq.py +++ b/GPT_SoVITS/module/core_vq.py @@ -34,6 +34,10 @@ from einops import rearrange, repeat import torch +try: + import torch_musa +except ImportError: + pass from torch import nn import torch.nn.functional as F from tqdm import tqdm diff --git a/GPT_SoVITS/module/mel_processing.py b/GPT_SoVITS/module/mel_processing.py index 503825ec..80e1bdd7 100644 --- a/GPT_SoVITS/module/mel_processing.py +++ b/GPT_SoVITS/module/mel_processing.py @@ -2,6 +2,10 @@ import os import random import torch +try: + import torch_musa +except ImportError: + pass from torch import nn import torch.nn.functional as F import torch.utils.data diff --git a/GPT_SoVITS/module/models.py b/GPT_SoVITS/module/models.py index 29676f43..9e2aeb1d 100644 --- a/GPT_SoVITS/module/models.py +++ b/GPT_SoVITS/module/models.py @@ -1,6 +1,10 @@ import copy import math import torch +try: + import torch_musa +except ImportError: + pass from torch import nn from torch.nn import functional as F diff --git a/GPT_SoVITS/module/modules.py b/GPT_SoVITS/module/modules.py index f4447455..06905fde 100644 --- a/GPT_SoVITS/module/modules.py +++ b/GPT_SoVITS/module/modules.py @@ -1,6 +1,10 @@ import math import numpy as np import torch +try: + import torch_musa +except ImportError: + pass from torch import nn from torch.nn import functional as F diff --git a/GPT_SoVITS/module/mrte_model.py b/GPT_SoVITS/module/mrte_model.py index b0cd242c..cf29c9e6 100644 --- a/GPT_SoVITS/module/mrte_model.py +++ b/GPT_SoVITS/module/mrte_model.py @@ -1,6 +1,10 @@ # This is Multi-reference timbre encoder import torch +try: + import torch_musa +except ImportError: + pass from torch import nn from torch.nn.utils import remove_weight_norm, weight_norm from module.attentions import MultiHeadAttention diff --git a/GPT_SoVITS/module/quantize.py b/GPT_SoVITS/module/quantize.py index f9a5c632..1d2354f2 100644 --- a/GPT_SoVITS/module/quantize.py +++ b/GPT_SoVITS/module/quantize.py @@ -11,6 +11,10 @@ import typing as tp import torch +try: + import torch_musa +except ImportError: + pass from torch import nn from module.core_vq import ResidualVectorQuantization diff --git a/GPT_SoVITS/module/transforms.py b/GPT_SoVITS/module/transforms.py index a11f799e..d859e49b 100644 --- a/GPT_SoVITS/module/transforms.py +++ b/GPT_SoVITS/module/transforms.py @@ -1,4 +1,8 @@ import torch +try: + import torch_musa +except ImportError: + pass from torch.nn import functional as F import numpy as np From 518a58e0ab5f33c6b1b86b863a34bf4902ed8ba0 Mon Sep 17 00:00:00 2001 From: KakaruHayate Date: Sat, 30 Mar 2024 22:50:54 +0800 Subject: [PATCH 06/10] Update README.md MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 只改了中文文档，因为这卡似乎只在大陆发售 --- docs/cn/README.md | 35 +++++++++++++++++++++++++++++++++++ 1 file changed, 35 insertions(+) diff --git a/docs/cn/README.md b/docs/cn/README.md index 5ff8a763..0bae2f3a 100644 --- a/docs/cn/README.md +++ b/docs/cn/README.md @@ -129,6 +129,41 @@ docker compose -f "docker-compose.yaml" up -d docker run --rm -it --gpus=all --env=is_half=False --volume=G:\GPT-SoVITS-DockerTest\output:/workspace/output --volume=G:\GPT-SoVITS-DockerTest\logs:/workspace/logs --volume=G:\GPT-SoVITS-DockerTest\SoVITS_weights:/workspace/SoVITS_weights --workdir=/workspace -p 9880:9880 -p 9871:9871 -p 9872:9872 -p 9873:9873 -p 9874:9874 --shm-size="16G" -d breakstring/gpt-sovits:xxxxx ``` +### 在摩尔线程显卡（MUSA）运行 + +只能运行在Ubuntu 20.04 LTS（内核版本5.4.X-5.15.X）下，非虚拟机，Intel CORE系列CPU + +1.前往[摩尔线程应用商店](https://developer.mthreads.com/sdk/download/musa?equipment=&os=Ubuntu&driverVersion=&version=)下载并按顺序安装`musa driver``musa_toolkit``mudnn``mccl` +2.前往[torch_musa](https://github.com/MooreThreads/torch_musa/releases/tag/v1.1.0)，根据你的显卡和python版本下载`torch``torch_musa`，将文件名`-linux_x86_64`后部分删除，使用以下命令安装 + +``` +pip install torch-2.0.0-cp39-cp39-linux_x86_64.whl +pip install torch_musa-1.1.0-cp39-cp39-linux_x86_64.whl +``` + +3.安装环境 + +``` +conda install -c conda-forge gcc +conda install -c conda-forge gxx +conda install ffmpeg cmake=3.18 ninja +``` + +之后你需要通过source安装torchaudio，因为摩尔线程官方并没有放出编译好的wheel +``` +git clone https://github.com/pytorch/audio +cd audio +USE_CUDA=0 python setup.py install +``` + +安装其他依赖 + +``` +pip install -r requirements.txt +``` + +4.运行`python webui.py` + ## 预训练模型 从 [GPT-SoVITS Models](https://huggingface.co/lj1995/GPT-SoVITS) 下载预训练模型，并将它们放置在 `GPT_SoVITS\pretrained_models` 中。 From e10fdf5f453b91cc46bf4acdb06c98462a45768d Mon Sep 17 00:00:00 2001 From: KakaruHayate Date: Sat, 30 Mar 2024 22:52:23 +0800 Subject: [PATCH 07/10] Update README.md --- docs/cn/README.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/docs/cn/README.md b/docs/cn/README.md index 0bae2f3a..52021828 100644 --- a/docs/cn/README.md +++ b/docs/cn/README.md @@ -133,8 +133,8 @@ docker run --rm -it --gpus=all --env=is_half=False --volume=G:\GPT-SoVITS-Docker 只能运行在Ubuntu 20.04 LTS（内核版本5.4.X-5.15.X）下，非虚拟机，Intel CORE系列CPU -1.前往[摩尔线程应用商店](https://developer.mthreads.com/sdk/download/musa?equipment=&os=Ubuntu&driverVersion=&version=)下载并按顺序安装`musa driver``musa_toolkit``mudnn``mccl` -2.前往[torch_musa](https://github.com/MooreThreads/torch_musa/releases/tag/v1.1.0)，根据你的显卡和python版本下载`torch``torch_musa`，将文件名`-linux_x86_64`后部分删除，使用以下命令安装 +1.前往[摩尔线程应用商店](https://developer.mthreads.com/sdk/download/musa?equipment=&os=Ubuntu&driverVersion=&version=)下载并按顺序安装`musa driver`、`musa_toolkit`、`mudnn`、`mccl` +2.前往[torch_musa](https://github.com/MooreThreads/torch_musa/releases/tag/v1.1.0)，根据你的显卡和python版本下载`torch`、`torch_musa`，将文件名`-linux_x86_64`后部分删除，使用以下命令安装 ``` pip install torch-2.0.0-cp39-cp39-linux_x86_64.whl From 5ddcf65b6055360b9c3ad2f1a8d9108c178c95ff Mon Sep 17 00:00:00 2001 From: KakaruHayate Date: Sat, 30 Mar 2024 22:53:31 +0800 Subject: [PATCH 08/10] Update README.md --- docs/cn/README.md | 9 +++++---- 1 file changed, 5 insertions(+), 4 deletions(-) diff --git a/docs/cn/README.md b/docs/cn/README.md index 52021828..f47bdc8b 100644 --- a/docs/cn/README.md +++ b/docs/cn/README.md @@ -133,15 +133,16 @@ docker run --rm -it --gpus=all --env=is_half=False --volume=G:\GPT-SoVITS-Docker 只能运行在Ubuntu 20.04 LTS（内核版本5.4.X-5.15.X）下，非虚拟机，Intel CORE系列CPU -1.前往[摩尔线程应用商店](https://developer.mthreads.com/sdk/download/musa?equipment=&os=Ubuntu&driverVersion=&version=)下载并按顺序安装`musa driver`、`musa_toolkit`、`mudnn`、`mccl` -2.前往[torch_musa](https://github.com/MooreThreads/torch_musa/releases/tag/v1.1.0)，根据你的显卡和python版本下载`torch`、`torch_musa`，将文件名`-linux_x86_64`后部分删除，使用以下命令安装 +1. 前往[摩尔线程应用商店](https://developer.mthreads.com/sdk/download/musa?equipment=&os=Ubuntu&driverVersion=&version=)下载并按顺序安装`musa driver`、`musa_toolkit`、`mudnn`、`mccl` + +2. 前往[torch_musa](https://github.com/MooreThreads/torch_musa/releases/tag/v1.1.0)，根据你的显卡和python版本下载`torch`、`torch_musa`，将文件名`-linux_x86_64`后部分删除，使用以下命令安装 ``` pip install torch-2.0.0-cp39-cp39-linux_x86_64.whl pip install torch_musa-1.1.0-cp39-cp39-linux_x86_64.whl ``` -3.安装环境 +3. 安装环境 ``` conda install -c conda-forge gcc @@ -162,7 +163,7 @@ USE_CUDA=0 python setup.py install pip install -r requirements.txt ``` -4.运行`python webui.py` +4. 运行`python webui.py` ## 预训练模型 From 4f87b534f4892426ddc9c4b25207b51a5ea684fd Mon Sep 17 00:00:00 2001 From: KakaruHayate Date: Sat, 30 Mar 2024 22:55:12 +0800 Subject: [PATCH 09/10] Update README.md --- docs/cn/README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/docs/cn/README.md b/docs/cn/README.md index f47bdc8b..1d0318f8 100644 --- a/docs/cn/README.md +++ b/docs/cn/README.md @@ -131,7 +131,7 @@ docker run --rm -it --gpus=all --env=is_half=False --volume=G:\GPT-SoVITS-Docker ### 在摩尔线程显卡（MUSA）运行 -只能运行在Ubuntu 20.04 LTS（内核版本5.4.X-5.15.X）下，非虚拟机，Intel CORE系列CPU +只能运行在Ubuntu 20.04 LTS（内核版本5.4.X-5.15.X）下，非虚拟机，Intel CORE系列CPU，**目前只支持推理** 1. 前往[摩尔线程应用商店](https://developer.mthreads.com/sdk/download/musa?equipment=&os=Ubuntu&driverVersion=&version=)下载并按顺序安装`musa driver`、`musa_toolkit`、`mudnn`、`mccl` From 019e3313646fdeaf4aabcf335d51de2ce9d764a2 Mon Sep 17 00:00:00 2001 From: KakaruHayate Date: Sun, 31 Mar 2024 11:47:37 +0800 Subject: [PATCH 10/10] API / CONFIG --- api.py | 8 ++++++-- config.py | 17 +++++++++++++++-- 2 files changed, 21 insertions(+), 4 deletions(-) diff --git a/api.py b/api.py index ea0e39d0..4f44d0b6 100644 --- a/api.py +++ b/api.py @@ -13,7 +13,7 @@ `-dt` - `默认参考音频文本` `-dl` - `默认参考音频语种, "中文","英文","日文","zh","en","ja"` -`-d` - `推理设备, "cuda","cpu"` +`-d` - `推理设备, "cuda","cpu","musa"` `-a` - `绑定地址, 默认"127.0.0.1"` `-p` - `绑定端口, 默认9880, 可在 config.py 中指定` `-fp` - `覆盖 config.py 使用全精度` @@ -124,6 +124,10 @@ import LangSegment from time import time as ttime import torch +try: + import torch_musa +except ImportError: + pass import librosa import soundfile as sf from fastapi import FastAPI, Request, HTTPException @@ -570,7 +574,7 @@ def handle(refer_wav_path, prompt_text, prompt_language, text, text_language, cu parser.add_argument("-dr", "--default_refer_path", type=str, default="", help="默认参考音频路径") parser.add_argument("-dt", "--default_refer_text", type=str, default="", help="默认参考音频文本") parser.add_argument("-dl", "--default_refer_language", type=str, default="", help="默认参考音频语种") -parser.add_argument("-d", "--device", type=str, default=g_config.infer_device, help="cuda / cpu") +parser.add_argument("-d", "--device", type=str, default=g_config.infer_device, help="cuda / cpu / MUSA ") parser.add_argument("-a", "--bind_addr", type=str, default="0.0.0.0", help="default: 0.0.0.0") parser.add_argument("-p", "--port", type=int, default=g_config.api_port, help="default: 9880") parser.add_argument("-fp", "--full_precision", action="store_true", default=False, help="覆盖config.is_half为False, 使用全精度") diff --git a/config.py b/config.py index 1f741285..d53789ec 100644 --- a/config.py +++ b/config.py @@ -17,10 +17,23 @@ exp_root = "logs" python_exec = sys.executable or "python" + +infer_device = "cpu" + +# 判断是否有摩尔线程显卡可用 +try: + import torch_musa + use_torch_musa = True +except ImportError: + use_torch_musa = False +if use_torch_musa: + if torch.musa.is_available(): + infer_device = "musa" + is_half=False + print("GPT-SoVITS running on MUSA!") + if torch.cuda.is_available(): infer_device = "cuda" -else: - infer_device = "cpu" webui_port_main = 9874 webui_port_uvr5 = 9873