Merge branch 'Stardust-minus:master' into master

README.md

@@ -3,3 +3,5 @@
 VITS2 Backbone with bert
 ## 成熟的旅行者/开拓者/舰长/博士/sensei/猎魔人/喵喵露/V应该参阅代码自己学习如何训练。
 ### 严禁将此项目用于一切违反《中华人民共和国宪法》，《中华人民共和国刑法》，《中华人民共和国治安管理处罚法》和《中华人民共和国民法典》之用途。
+#### Video:https://www.bilibili.com/video/BV1hp4y1K78E   
+#### Demo:https://www.bilibili.com/video/BV1TF411k78w

attentions.py

@@ -43,11 +43,11 @@ class Encoder(nn.Module):
     self.kernel_size = kernel_size
     self.p_dropout = p_dropout
     self.window_size = window_size
-    if isflow:
+    #if isflow:
-      cond_layer = torch.nn.Conv1d(256, 2*hidden_channels*n_layers, 1)
+    #  cond_layer = torch.nn.Conv1d(256, 2*hidden_channels*n_layers, 1)
-      self.cond_pre = torch.nn.Conv1d(hidden_channels, 2*hidden_channels, 1)
+    #  self.cond_pre = torch.nn.Conv1d(hidden_channels, 2*hidden_channels, 1)
-      self.cond_layer = weight_norm(cond_layer, name='weight')
+    #  self.cond_layer = weight_norm(cond_layer, name='weight')
-      self.gin_channels = 256
+    #  self.gin_channels = 256
     self.cond_layer_idx = self.n_layers
     if 'gin_channels' in kwargs:
       self.gin_channels = kwargs['gin_channels']

bert_gen.py

@@ -1,11 +1,11 @@
 import torch
 from torch.utils.data import DataLoader
-
+from multiprocessing import Pool
 import commons
 import utils
 from data_utils import TextAudioSpeakerLoader, TextAudioSpeakerCollate
 from tqdm import tqdm
-from multiprocessing import Pool
+import warnings
 
 from text import cleaned_text_to_sequence, get_bert
 
@@ -40,14 +40,14 @@ def process_line(line):
         torch.save(bert, bert_path)
 
 
-lines = []
+if __name__ == '__main__':
-with open(hps.data.training_files) as f:
+    lines = []
-    lines.extend(f.readlines())
+    with open(hps.data.training_files, encoding='utf-8' ) as f:
+        lines.extend(f.readlines())
 
-with open(hps.data.validation_files) as f:
+    with open(hps.data.validation_files, encoding='utf-8' ) as f:
-    lines.extend(f.readlines())
+        lines.extend(f.readlines())
 
-
+    with Pool(processes=12) as pool: #A100 40GB suitable config,if coom,please decrease the processess number.
-with Pool(processes=12) as pool: #A100 suitable config,if coom,please decrease the processess number.
+        for _ in tqdm(pool.imap_unordered(process_line, lines)):
-    for _ in tqdm(pool.imap_unordered(process_line, lines)):
+            pass
-        pass

models.py

@@ -15,8 +15,6 @@ from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
 from commons import init_weights, get_padding
 from text import symbols, num_tones, num_languages
 class DurationDiscriminator(nn.Module): #vits2
-  # TODO : not using "spk conditioning" for now according to the paper.
-  # Can be a better discriminator if we use it.
   def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, gin_channels=0):
     super().__init__()
 
@@ -34,11 +32,11 @@ class DurationDiscriminator(nn.Module): #vits2
     self.dur_proj = nn.Conv1d(1, filter_channels, 1)
 
     self.pre_out_conv_1 = nn.Conv1d(2*filter_channels, filter_channels, kernel_size, padding=kernel_size//2)
-    self.pre_out_norm_1 = modules.LayerNorm(filter_channels)
+    #self.pre_out_norm_1 = modules.LayerNorm(filter_channels)
     self.pre_out_conv_2 = nn.Conv1d(filter_channels, filter_channels, kernel_size, padding=kernel_size//2)
-    self.pre_out_norm_2 = modules.LayerNorm(filter_channels)
+    #self.pre_out_norm_2 = modules.LayerNorm(filter_channels)
 
-    # if gin_channels != 0:
+    #if gin_channels != 0:
     #   self.cond = nn.Conv1d(gin_channels, in_channels, 1)
 
     self.output_layer = nn.Sequential(
@@ -64,7 +62,7 @@ class DurationDiscriminator(nn.Module): #vits2
 
   def forward(self, x, x_mask, dur_r, dur_hat, g=None):
     x = torch.detach(x)
-    # if g is not None:
+    #if g is not None:
     #   g = torch.detach(g)
     #   x = x + self.cond(g)
     x = self.conv_1(x * x_mask)

preprocess_text.py

@@ -23,8 +23,8 @@ if 1 in stage:
                 f.write('{}|{}|{}|{}|{}|{}|{}\n'.format(utt, spk, language, norm_text, ' '.join(phones),
                                                      " ".join([str(i) for i in tones]),
                                                      " ".join([str(i) for i in word2ph])))
-            except:
+            except Exception as error :
-                print("err!", utt)
+                print("err!", utt, error)
 
 if 2 in stage:
     spk_utt_map = defaultdict(list)

train_ms.py

@@ -78,7 +78,8 @@ def run(rank, n_gpus, hps):
         shuffle=True)
     collate_fn = TextAudioSpeakerCollate()
     train_loader = DataLoader(train_dataset, num_workers=24, shuffle=False, pin_memory=True,
-                              collate_fn=collate_fn, batch_sampler=train_sampler, persistent_workers=True,prefetch_factor=4)
+                              collate_fn=collate_fn, batch_sampler=train_sampler,
+                              persistent_workers=True,prefetch_factor=4)  #256G Memory suitable loader.
     if rank == 0:
         eval_dataset = TextAudioSpeakerLoader(hps.data.validation_files, hps.data)
         eval_loader = DataLoader(eval_dataset, num_workers=0, shuffle=False,
@@ -146,8 +147,8 @@ def run(rank, n_gpus, hps):
         eps=hps.train.eps)
     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], find_unused_parameters=True)
-    net_d = DDP(net_d, device_ids=[rank],find_unused_parameters=True)
+    net_d = DDP(net_d, device_ids=[rank], find_unused_parameters=True)
     if net_dur_disc is not None:
         net_dur_disc = DDP(net_dur_disc, device_ids=[rank], find_unused_parameters=True)