update

models.py

@@ -515,7 +515,7 @@ class SynthesizerTrn(nn.Module):
         if n_speakers > 1:
             self.emb_g = nn.Embedding(n_speakers, gin_channels)
         else:
-            self.ref_enc = ReferenceEncoder(spec_channels)
+            self.ref_enc = ReferenceEncoder(spec_channels, gin_channels)
 
     def forward(self, x, x_lengths, y, y_lengths, sid, tone, language, bert):
 

preprocess_text.py

@@ -10,7 +10,7 @@ transcription_path = 'filelists/aishell.list'
 train_path = 'filelists/train.list'
 val_path = 'filelists/val.list'
 config_path = "configs/config.json"
-val_per_spk = 2
+val_per_spk = 4
 max_val_total = 8
 
 if 1 in stage:

spec_gen.py

@@ -2,11 +2,11 @@ import utils
 from data_utils import TextAudioSpeakerLoader
 from tqdm import tqdm
 
-config_path = 'configs/config.json'
+config_path = 'configs/fzh.json'
 hps = utils.get_hparams_from_file(config_path)
 
-train_dataset = TextAudioSpeakerLoader("filelists/train.list", hps.data)
-eval_dataset = TextAudioSpeakerLoader("filelists/val.list", hps.data)
+train_dataset = TextAudioSpeakerLoader(hps.data.training_files, hps.data)
+eval_dataset = TextAudioSpeakerLoader(hps.data.validation_files, hps.data)
 
 for _ in tqdm(train_dataset):
     pass

train_ms.py

@@ -279,7 +279,7 @@ def evaluate(hps, generator, eval_loader, writer_eval):
             bert = bert.cuda()
             tone = tone.cuda()
             language = language.cuda()
-            for use_sdp in [False, True]:
+            for use_sdp in [True, False]:
                 y_hat, attn, mask, *_ = generator.module.infer(x, x_lengths, speakers, tone, language, bert, y=spec, max_len=1000, sdp_ratio=0.0 if not use_sdp else 1.0)
                 y_hat_lengths = mask.sum([1, 2]).long() * hps.data.hop_length