sync emo branch (#159)

* Update README.md

* 更新 bert_models.json

* fix

* Update data_utils.py

* Update infer.py

* performance improve

* Feat: support auto split in webui (#158)

* Feat: support auto split in webui

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

---------

Co-authored-by: Sora <atri@suzakuintsubaki.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

README.md

@@ -1,7 +1,9 @@
 # Bert-VITS2
 
 VITS2 Backbone with bert
-
+# 紧急通知
+我们在2.0版本中发现了重大bug，该bug导致日文和英文bert被置0后训练，即失去bert效果。
+我们将重炼2.0版本底模，已经开炉的建议关炉静候。
 ## 请注意，本项目核心思路来源于[anyvoiceai/MassTTS](https://github.com/anyvoiceai/MassTTS) 一个非常好的tts项目
 ## MassTTS的演示demo为[ai版峰哥锐评峰哥本人,并找回了在金三角失落的腰子](https://www.bilibili.com/video/BV1w24y1c7z9)
 

bert/bert_models.json

@@ -1,7 +1,7 @@
 {
     "deberta-v2-large-japanese": {
         "repo_id": "ku-nlp/deberta-v2-large-japanese",
-        "files": ["spm.model", "pytorch_model.bin"]
+        "files": ["pytorch_model.bin"]
     },
     "chinese-roberta-wwm-ext-large": {
         "repo_id": "hfl/chinese-roberta-wwm-ext-large",

data_utils.py

@@ -145,38 +145,24 @@ class TextAudioSpeakerLoader(torch.utils.data.Dataset):
             word2ph[0] += 1
         bert_path = wav_path.replace(".wav", ".bert.pt")
         try:
-            bert = torch.load(bert_path)
+            bert_ori = torch.load(bert_path)
-            assert bert.shape[-1] == len(phone)
+            assert bert_ori.shape[-1] == len(phone)
         except Exception as e:
             logger.warn("Bert load Failed")
             logger.warn(e)
 
         if language_str == "ZH":
-            bert = bert
+            bert = bert_ori
             ja_bert = torch.zeros(1024, len(phone))
             en_bert = torch.zeros(1024, len(phone))
         elif language_str == "JP":
             bert = torch.zeros(1024, len(phone))
-            ja_bert = bert
+            ja_bert = bert_ori
             en_bert = torch.zeros(1024, len(phone))
         elif language_str == "EN":
             bert = torch.zeros(1024, len(phone))
             ja_bert = torch.zeros(1024, len(phone))
-            en_bert = bert
+            en_bert = bert_ori
-        assert bert.shape[-1] == len(phone), (
-            bert.shape,
-            len(phone),
-            sum(word2ph),
-            p1,
-            p2,
-            t1,
-            t2,
-            pold,
-            pold2,
-            word2ph,
-            text,
-            w2pho,
-        )
         phone = torch.LongTensor(phone)
         tone = torch.LongTensor(tone)
         language = torch.LongTensor(language)

infer.py

@@ -85,22 +85,22 @@ def get_text(text, language_str, hps, device):
         for i in range(len(word2ph)):
             word2ph[i] = word2ph[i] * 2
         word2ph[0] += 1
-    bert = get_bert(norm_text, word2ph, language_str, device)
+    bert_ori = get_bert(norm_text, word2ph, language_str, device)
     del word2ph
-    assert bert.shape[-1] == len(phone), phone
+    assert bert_ori.shape[-1] == len(phone), phone
 
     if language_str == "ZH":
-        bert = bert
+        bert = bert_ori
         ja_bert = torch.zeros(1024, len(phone))
         en_bert = torch.zeros(1024, len(phone))
     elif language_str == "JP":
         bert = torch.zeros(1024, len(phone))
-        ja_bert = bert
+        ja_bert = bert_ori
         en_bert = torch.zeros(1024, len(phone))
     elif language_str == "EN":
         bert = torch.zeros(1024, len(phone))
         ja_bert = torch.zeros(1024, len(phone))
-        en_bert = bert
+        en_bert = bert_ori
     else:
         raise ValueError("language_str should be ZH, JP or EN")
 

server_fastapi.py

@@ -5,6 +5,7 @@ import logging
 import gc
 import random
 
+import gradio
 import numpy as np
 import utils
 from fastapi import FastAPI, Query, Request
@@ -245,6 +246,7 @@ if __name__ == "__main__":
                     )
                 audios.append(np.zeros((int)(44100 * 0.3)))
                 audio = np.concatenate(audios)
+                audio = gradio.processing_utils.convert_to_16_bit_wav(audio)
         wavContent = BytesIO()
         wavfile.write(
             wavContent, loaded_models.models[model_id].hps.data.sampling_rate, audio

tools/log.py

@@ -9,6 +9,8 @@ import sys
 logger.remove()
 
 # 自定义格式并添加到标准输出
-log_format = "<g>{time:MM-DD HH:mm:ss}</g> [<lvl>{level}</lvl>] | {message}"
+log_format = (
+    "<g>{time:MM-DD HH:mm:ss}</g> [<lvl>{level}</lvl>] | {file}:{line} | {message}"
+)
 
 logger.add(sys.stdout, format=log_format)

train_ms.py

@@ -206,8 +206,8 @@ def run():
         )
     else:
         optim_dur_disc = None
-    net_g = DDP(net_g, device_ids=[rank], find_unused_parameters=True)
+    net_g = DDP(net_g, device_ids=[rank])
-    net_d = DDP(net_d, device_ids=[rank], find_unused_parameters=True)
+    net_d = DDP(net_d, device_ids=[rank])
     dur_resume_lr = None
     if net_dur_disc is not None:
         net_dur_disc = DDP(net_dur_disc, device_ids=[rank], find_unused_parameters=True)