Merge branch 'RVC-Boss:main' into main

GPT_SoVITS/inference_webui.py

@@ -12,6 +12,8 @@ bert_path = os.environ.get(
 )
 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"))
@@ -115,7 +117,6 @@ vq_model.eval()
 print(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:
@@ -149,13 +150,21 @@ def get_tts_wav(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")
+    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)  # 派蒙
+        wav16k, sr = librosa.load(ref_wav_path, sr=16000)
         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(
@@ -170,10 +179,6 @@ def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_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:
         # 解决输入目标文本的空行导致报错的问题
         if (len(text.strip()) == 0):

GPT_SoVITS/module/data_utils.py

@@ -1,6 +1,8 @@
-import time, logging
+import time
+import logging
 import os
-import random, traceback
+import random
+import traceback
 import numpy as np
 import torch
 import torch.utils.data
@@ -12,15 +14,12 @@ from text import cleaned_text_to_sequence
 from utils import load_wav_to_torch, load_filepaths_and_text
 import torch.nn.functional as F
 from functools import lru_cache
-import torch
 import requests
 from scipy.io import wavfile
 from io import BytesIO
-
-# from config import exp_dir
 from my_utils import load_audio
 
-
+# ZeroDivisionError fixed by Tybost (https://github.com/RVC-Boss/GPT-SoVITS/issues/79)
 class TextAudioSpeakerLoader(torch.utils.data.Dataset):
     """
     1) loads audio, speaker_id, text pairs
@@ -44,7 +43,7 @@ class TextAudioSpeakerLoader(torch.utils.data.Dataset):
 
         for line in lines:
             tmp = line.split("\t")
-            if len(tmp) != 4:
+            if (len(tmp) != 4):
                 continue
             self.phoneme_data[tmp[0]] = [tmp[1]]
 
@@ -52,7 +51,7 @@ class TextAudioSpeakerLoader(torch.utils.data.Dataset):
         tmp = self.audiopaths_sid_text
         leng = len(tmp)
         min_num = 100
-        if leng < min_num:
+        if (leng < min_num):
             self.audiopaths_sid_text = []
             for _ in range(max(2, int(min_num / leng))):
                 self.audiopaths_sid_text += tmp
@@ -77,20 +76,28 @@ class TextAudioSpeakerLoader(torch.utils.data.Dataset):
         for audiopath in tqdm(self.audiopaths_sid_text):
             try:
                 phoneme = self.phoneme_data[audiopath][0]
-                phoneme = phoneme.split(" ")
+                phoneme = phoneme.split(' ')
                 phoneme_ids = cleaned_text_to_sequence(phoneme)
             except Exception:
                 print(f"{audiopath} not in self.phoneme_data !")
                 skipped_phone += 1
                 continue
+
             size = os.path.getsize("%s/%s" % (self.path5, audiopath))
             duration = size / self.sampling_rate / 2
+
+            if duration == 0:
+                print(f"Zero duration for {audiopath}, skipping...")
+                skipped_dur += 1
+                continue
+
             if 54 > duration > 0.6 or self.val:
                 audiopaths_sid_text_new.append([audiopath, phoneme_ids])
                 lengths.append(size // (2 * self.hop_length))
             else:
                 skipped_dur += 1
                 continue
+
         print("skipped_phone: ", skipped_phone, ", skipped_dur: ", skipped_dur)
         print("total left: ", len(audiopaths_sid_text_new))
         assert len(audiopaths_sid_text_new) > 1  # 至少能凑够batch size，这里todo
@@ -103,10 +110,8 @@ class TextAudioSpeakerLoader(torch.utils.data.Dataset):
         try:
             spec, wav = self.get_audio("%s/%s" % (self.path5, audiopath))
             with torch.no_grad():
-                ssl = torch.load(
+                ssl = torch.load("%s/%s.pt" % (self.path4, audiopath), map_location="cpu")
-                    "%s/%s.pt" % (self.path4, audiopath), map_location="cpu"
+                if (ssl.shape[-1] != spec.shape[-1]):
-                )
-                if ssl.shape[-1] != spec.shape[-1]:
                     typee = ssl.dtype
                     ssl = F.pad(ssl.float(), (0, 1), mode="replicate").to(typee)
                 ssl.requires_grad = False
@@ -117,25 +122,15 @@ class TextAudioSpeakerLoader(torch.utils.data.Dataset):
             ssl = torch.zeros(1, 768, 100)
             text = text[-1:]
             print("load audio or ssl error!!!!!!", audiopath)
-        # print(ssl.requires_grad,spec.requires_grad,wav.requires_grad,text.requires_grad)
         return (ssl, spec, wav, text)
 
     def get_audio(self, filename):
-        audio_array = load_audio(
+        audio_array = load_audio(filename, self.sampling_rate)  # load_audio的方法是已经归一化到-1~1之间的，不用再/32768
-            filename, self.sampling_rate
-        )  # load_audio的方法是已经归一化到-1~1之间的，不用再/32768
-        # print(filename,audio_array.max(),audio_array.min(),audio_array.mean())
         audio = torch.FloatTensor(audio_array)  # /32768
         audio_norm = audio
         audio_norm = audio_norm.unsqueeze(0)
-        spec = spectrogram_torch(
+        spec = spectrogram_torch(audio_norm, self.filter_length, self.sampling_rate, self.hop_length, self.win_length,
-            audio_norm,
+                                  center=False)
-            self.filter_length,
-            self.sampling_rate,
-            self.hop_length,
-            self.win_length,
-            center=False,
-        )
         spec = torch.squeeze(spec, 0)
         return spec, audio_norm
 
@@ -152,14 +147,11 @@ class TextAudioSpeakerLoader(torch.utils.data.Dataset):
 
     def random_slice(self, ssl, wav, mel):
         assert abs(ssl.shape[-1] - wav.shape[-1] // self.hop_length) < 3, (
-            "first",
+        "first", ssl.shape, wav.shape)
-            ssl.shape,
-            wav.shape,
-        )
 
         len_mel = mel.shape[1]
         if self.val:
-            reference_mel = mel[:, : len_mel // 3]
+            reference_mel = mel[:, :len_mel // 3]
             return reference_mel, ssl, wav, mel
         dir = random.randint(0, 1)
         sep_point = random.randint(int(len_mel // 3), int(len_mel // 3 * 2))
@@ -167,29 +159,22 @@ class TextAudioSpeakerLoader(torch.utils.data.Dataset):
         if dir == 0:
             reference_mel = mel[:, :sep_point]
             ssl = ssl[:, :, sep_point:]
-            wav2 = wav[:, sep_point * self.hop_length :]
+            wav2 = wav[:, sep_point * self.hop_length:]
             mel = mel[:, sep_point:]
         else:
             reference_mel = mel[:, sep_point:]
             ssl = ssl[:, :, :sep_point]
-            wav2 = wav[:, : sep_point * self.hop_length]
+            wav2 = wav[:, :sep_point * self.hop_length]
             mel = mel[:, :sep_point]
 
         assert abs(ssl.shape[-1] - wav2.shape[-1] // self.hop_length) < 3, (
-            ssl.shape,
+        ssl.shape, wav.shape, wav2.shape, mel.shape, sep_point, self.hop_length, sep_point * self.hop_length, dir)
-            wav.shape,
-            wav2.shape,
-            mel.shape,
-            sep_point,
-            self.hop_length,
-            sep_point * self.hop_length,
-            dir,
-        )
         return reference_mel, ssl, wav2, mel
 
 
-class TextAudioSpeakerCollate:
+class TextAudioSpeakerCollate():
-    """Zero-pads model inputs and targets"""
+    """ Zero-pads model inputs and targets
+    """
 
     def __init__(self, return_ids=False):
         self.return_ids = return_ids
@@ -202,8 +187,8 @@ class TextAudioSpeakerCollate:
         """
         # Right zero-pad all one-hot text sequences to max input length
         _, ids_sorted_decreasing = torch.sort(
-            torch.LongTensor([x[1].size(1) for x in batch]), dim=0, descending=True
+            torch.LongTensor([x[1].size(1) for x in batch]),
-        )
+            dim=0, descending=True)
 
         max_ssl_len = max([x[0].size(2) for x in batch])
         max_ssl_len = int(2 * ((max_ssl_len // 2) + 1))
@@ -231,31 +216,22 @@ class TextAudioSpeakerCollate:
             row = batch[ids_sorted_decreasing[i]]
 
             ssl = row[0]
-            ssl_padded[i, :, : ssl.size(2)] = ssl[0, :, :]
+            ssl_padded[i, :, :ssl.size(2)] = ssl[0, :, :]
             ssl_lengths[i] = ssl.size(2)
 
             spec = row[1]
-            spec_padded[i, :, : spec.size(1)] = spec
+            spec_padded[i, :, :spec.size(1)] = spec
             spec_lengths[i] = spec.size(1)
 
             wav = row[2]
-            wav_padded[i, :, : wav.size(1)] = wav
+            wav_padded[i, :, :wav.size(1)] = wav
             wav_lengths[i] = wav.size(1)
 
             text = row[3]
-            text_padded[i, : text.size(0)] = text
+            text_padded[i, :text.size(0)] = text
             text_lengths[i] = text.size(0)
 
-        return (
+        return ssl_padded, ssl_lengths, spec_padded, spec_lengths, wav_padded, wav_lengths, text_padded, text_lengths
-            ssl_padded,
-            ssl_lengths,
-            spec_padded,
-            spec_lengths,
-            wav_padded,
-            wav_lengths,
-            text_padded,
-            text_lengths,
-        )
 
 
 class DistributedBucketSampler(torch.utils.data.distributed.DistributedSampler):
@@ -268,18 +244,9 @@ class DistributedBucketSampler(torch.utils.data.distributed.DistributedSampler):
     Ex) boundaries = [b1, b2, b3] -> any x s.t. length(x) <= b1 or length(x) > b3 are discarded.
     """
 
-    def __init__(
+    def __init__(self, dataset, batch_size, boundaries, num_replicas=None, rank=None, shuffle=True):
-        self,
-        dataset,
-        batch_size,
-        boundaries,
-        num_replicas=None,
-        rank=None,
-        shuffle=True,
-    ):
         super().__init__(dataset, num_replicas=num_replicas, rank=rank, shuffle=shuffle)
         self.lengths = dataset.lengths
-        # print(233333333333333,self.lengths,dir(dataset))
         self.batch_size = batch_size
         self.boundaries = boundaries
 
@@ -295,24 +262,22 @@ class DistributedBucketSampler(torch.utils.data.distributed.DistributedSampler):
             if idx_bucket != -1:
                 buckets[idx_bucket].append(i)
 
-        for i in range(len(buckets) - 1, 0, -1):
+        i = len(buckets) - 1
-            # for i in range(len(buckets) - 1, -1, -1):
+        while i >= 0:
             if len(buckets[i]) == 0:
                 buckets.pop(i)
                 self.boundaries.pop(i + 1)
+            i -= 1
 
         num_samples_per_bucket = []
         for i in range(len(buckets)):
             len_bucket = len(buckets[i])
             total_batch_size = self.num_replicas * self.batch_size
-            rem = (
+            rem = (total_batch_size - (len_bucket % total_batch_size)) % total_batch_size
-                total_batch_size - (len_bucket % total_batch_size)
-            ) % total_batch_size
             num_samples_per_bucket.append(len_bucket + rem)
         return buckets, num_samples_per_bucket
 
     def __iter__(self):
-        # deterministically shuffle based on epoch
         g = torch.Generator()
         g.manual_seed(self.epoch)
 
@@ -331,25 +296,13 @@ class DistributedBucketSampler(torch.utils.data.distributed.DistributedSampler):
             ids_bucket = indices[i]
             num_samples_bucket = self.num_samples_per_bucket[i]
 
-            # add extra samples to make it evenly divisible
             rem = num_samples_bucket - len_bucket
-            ids_bucket = (
+            ids_bucket = ids_bucket + ids_bucket * (rem // len_bucket) + ids_bucket[:(rem % len_bucket)]
-                ids_bucket
-                + ids_bucket * (rem // len_bucket)
-                + ids_bucket[: (rem % len_bucket)]
-            )
 
-            # subsample
+            ids_bucket = ids_bucket[self.rank::self.num_replicas]
-            ids_bucket = ids_bucket[self.rank :: self.num_replicas]
 
-            # batching
             for j in range(len(ids_bucket) // self.batch_size):
-                batch = [
+                batch = [bucket[idx] for idx in ids_bucket[j * self.batch_size:(j + 1) * self.batch_size]]
-                    bucket[idx]
-                    for idx in ids_bucket[
-                        j * self.batch_size : (j + 1) * self.batch_size
-                    ]
-                ]
                 batches.append(batch)
 
         if self.shuffle:
@@ -376,4 +329,4 @@ class DistributedBucketSampler(torch.utils.data.distributed.DistributedSampler):
             return -1
 
     def __len__(self):
-        return self.num_samples // self.batch_size
+        return self.num_samples // self.batch_size

config.py

@@ -5,6 +5,7 @@ import sys
 sovits_path = ""
 gpt_path = ""
 is_half = True
+is_share=False
 
 cnhubert_path = "GPT_SoVITS/pretrained_models/chinese-hubert-base"
 bert_path = "GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large"

docs/cn/Changelog_CN.md

@@ -0,0 +1,25 @@
+### 20240121更新
+
+1-config添加is_share，诸如colab等场景可以将此改为True，来使得webui映射到公网
+
+2-WebUI添加英文系统英文翻译适配
+
+3-cmd-asr自动判断是否已自带damo模型，如不在默认目录上将从modelscope自带下载
+
+4-[SoVITS训练报错ZeroDivisionError](https://github.com/RVC-Boss/GPT-SoVITS/issues/79) 尝试修复（过滤长度0的样本等）
+
+5-清理TEMP文件夹缓存音频等文件
+
+6-在参考音频结尾留空0.3s，削弱合成音频包含参考音频结尾的问题
+
+待修复：
+
+1-过短输出文件返回重复参考音频的问题
+
+2-batch size超过条数导致微调有问题
+
+3-hubert提取在half下出现nan概率更高的问题
+
+高优：
+
+支持英文日文训练

docs/cn/README.md

@@ -1,7 +1,7 @@
 <div align="center">
 
 <h1>GPT-SoVITS-WebUI</h1>
-少样本强大的声音转换与文本到语音网络界面。<br><br>
+强大的少样本语音转换与语音合成Web用户界面。<br><br>
 
 [![madewithlove](https://img.shields.io/badge/made_with-%E2%9D%A4-red?style=for-the-badge&labelColor=orange
 )](https://github.com/RVC-Boss/GPT-SoVITS)

requirements.txt

@@ -20,3 +20,4 @@ transformers
 chardet
 PyYAML
 psutil
+jieba

tools/damo_asr/cmd-asr.py

@@ -6,11 +6,18 @@ import sys,os,traceback
 dir=sys.argv[1]
 # opt_name=dir.split("\\")[-1].split("/")[-1]
 opt_name=os.path.basename(dir)
+
+path_asr='tools/damo_asr/models/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch'
+path_vad='tools/damo_asr/models/speech_fsmn_vad_zh-cn-16k-common-pytorch'
+path_punc='tools/damo_asr/models/punc_ct-transformer_zh-cn-common-vocab272727-pytorch'
+path_asr=path_asr if os.path.exists(path_asr)else "damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
+path_vad=path_vad if os.path.exists(path_vad)else "damo/speech_fsmn_vad_zh-cn-16k-common-pytorch"
+path_punc=path_punc if os.path.exists(path_punc)else "damo/punc_ct-transformer_zh-cn-common-vocab272727-pytorch"
 inference_pipeline = pipeline(
     task=Tasks.auto_speech_recognition,
-    model='tools/damo_asr/models/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch',
+    model=path_asr,
-    vad_model='tools/damo_asr/models/speech_fsmn_vad_zh-cn-16k-common-pytorch',
+    vad_model=path_vad,
-    punc_model='tools/damo_asr/models/punc_ct-transformer_zh-cn-common-vocab272727-pytorch',
+    punc_model=path_punc,
 )
 
 opt=[]

tools/subfix_webui.py

@@ -79,6 +79,7 @@ def b_change_index(index, batch):
 
 
 def b_next_index(index, batch):
+    b_save_file()
     if (index + batch) <= g_max_json_index:
         return index + batch , *b_change_index(index + batch, batch)
     else:
@@ -86,6 +87,7 @@ def b_next_index(index, batch):
 
 
 def b_previous_index(index, batch):
+    b_save_file()
     if (index - batch) >= 0:
         return index - batch , *b_change_index(index - batch, batch)
     else:
@@ -294,6 +296,7 @@ def set_global(load_json, load_list, json_key_text, json_key_path, batch):
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(description='Process some integers.')
     parser.add_argument('--load_json', default="None", help='source file, like demo.json')
+    parser.add_argument('--is_share', default="False", help='whether webui is_share=True')
     parser.add_argument('--load_list', default="None", help='source file, like demo.list')
     parser.add_argument('--webui_port_subfix', default=9871, help='source file, like demo.list')
     parser.add_argument('--json_key_text', default="text", help='the text key name in json, Default: text')
@@ -488,5 +491,6 @@ if __name__ == "__main__":
         server_name="0.0.0.0",
         inbrowser=True,
         quiet=True,
+        share=eval(args.is_share),
         server_port=int(args.webui_port_subfix)
     )

tools/uvr5/webui.py

@@ -19,7 +19,8 @@ for name in os.listdir(weight_uvr5_root):
 
 device=sys.argv[1]
 is_half=sys.argv[2]
-
+webui_port_uvr5=int(sys.argv[3])
+is_share=eval(sys.argv[4])
 
 def uvr(model_name, inp_root, save_root_vocal, paths, save_root_ins, agg, format0):
     infos = []
@@ -179,6 +180,7 @@ with gr.Blocks(title="RVC WebUI") as app:
 app.queue(concurrency_count=511, max_size=1022).launch(
     server_name="0.0.0.0",
     inbrowser=True,
-    server_port=9873,
+    share=is_share,
+    server_port=webui_port_uvr5,
     quiet=True,
 )

webui.py

@@ -1,7 +1,7 @@
 import json,yaml,warnings,torch
 import platform
 import psutil
-import os
+import os,shutil
 import signal
 from tools import my_utils
 
@@ -12,6 +12,12 @@ now_dir = os.getcwd()
 tmp = os.path.join(now_dir, "TEMP")
 os.makedirs(tmp, exist_ok=True)
 os.environ["TEMP"] = tmp
+if(os.path.exists(tmp)):
+    for name in os.listdir(tmp):
+        if(name=="jieba.cache"):continue
+        path="%s/%s"%(tmp,name)
+        delete=os.remove if os.path.isfile(path) else shutil.rmtree
+        delete(path)
 import site
 site_packages_roots = []
 for path in site.getsitepackages():
@@ -34,7 +40,7 @@ import pdb
 import gradio as gr
 from subprocess import Popen
 import signal
-from config import python_exec,infer_device,is_half,exp_root,webui_port_main,webui_port_infer_tts,webui_port_uvr5,webui_port_subfix
+from config import python_exec,infer_device,is_half,exp_root,webui_port_main,webui_port_infer_tts,webui_port_uvr5,webui_port_subfix,is_share
 from tools.i18n.i18n import I18nAuto
 i18n = I18nAuto()
 from scipy.io import wavfile
@@ -120,7 +126,7 @@ def kill_process(pid):
 def change_label(if_label,path_list):
     global p_label
     if(if_label==True and p_label==None):
-        cmd = '"%s" tools/subfix_webui.py --load_list "%s" --webui_port %s'%(python_exec,path_list,webui_port_subfix)
+        cmd = '"%s" tools/subfix_webui.py --load_list "%s" --webui_port %s --is_share %s'%(python_exec,path_list,webui_port_subfix,is_share)
         yield i18n("打标工具WebUI已开启")
         print(cmd)
         p_label = Popen(cmd, shell=True)
@@ -132,7 +138,7 @@ def change_label(if_label,path_list):
 def change_uvr5(if_uvr5):
     global p_uvr5
     if(if_uvr5==True and p_uvr5==None):
-        cmd = '"%s" tools/uvr5/webui.py "%s" %s %s'%(python_exec,infer_device,is_half,webui_port_uvr5)
+        cmd = '"%s" tools/uvr5/webui.py "%s" %s %s %s'%(python_exec,infer_device,is_half,webui_port_uvr5,is_share)
         yield i18n("UVR5已开启")
         print(cmd)
         p_uvr5 = Popen(cmd, shell=True)
@@ -151,6 +157,7 @@ def change_tts_inference(if_tts,bert_path,cnhubert_base_path,gpu_number,gpt_path
         os.environ["_CUDA_VISIBLE_DEVICES"]=gpu_number
         os.environ["is_half"]=str(is_half)
         os.environ["infer_ttswebui"]=str(webui_port_infer_tts)
+        os.environ["is_share"]=str(is_share)
         cmd = '"%s" GPT_SoVITS/inference_webui.py'%(python_exec)
         yield i18n("TTS推理进程已开启")
         print(cmd)
@@ -659,7 +666,7 @@ with gr.Blocks(title="GPT-SoVITS WebUI") as app:
                         label=i18n("*训练集音频文件目录"),
                         # value=r"D:\RVC1006\GPT-SoVITS\raw\xxx",
                         interactive=True,
-                        placeholder=i18n("训练集音频文件目录 拼接 list文件里波形对应的文件名。")
+                        placeholder=i18n("训练集音频文件目录-拼接-list文件里波形对应的文件名（不是全路径）。")
                     )
                 gr.Markdown(value=i18n("1Aa-文本内容"))
                 with gr.Row():
@@ -740,7 +747,7 @@ with gr.Blocks(title="GPT-SoVITS WebUI") as app:
     app.queue(concurrency_count=511, max_size=1022).launch(
         server_name="0.0.0.0",
         inbrowser=True,
-        share=True,
+        share=is_share,
         server_port=webui_port_main,
         quiet=True,
     )