thweb finetune

config.json

@@ -12,7 +12,7 @@
     "text_cleaner": "english_cleaners",
 
     "epochs": 50,
-    "lr": 0.002,
+    "lr": 0.004,
     "warmup_steps": 4000,
     "batch_size": 32,
     "eval_batch_size":32,
@@ -23,10 +23,10 @@
     "griffin_lim_iters": 60,
     "power": 1.2,
 
-    "dataset": "LJSpeech",
+    "dataset": "TWEB",
-    "meta_file_train": "metadata_train.csv",
+    "meta_file_train": "transcript_train.txt",
-    "meta_file_val": "metadata_val.csv",
+    "meta_file_val": "transcript_val.txt",
-    "data_path": "/data/shared/KeithIto/LJSpeech-1.0/",
+    "data_path": "/data/shared/BibleSpeech/",
     "min_seq_len": 0, 
     "num_loader_workers": 8,
 

layers/tacotron.py

@@ -324,4 +324,4 @@ class Decoder(nn.Module):
 
 def is_end_of_frames(output, alignment, eps=0.01):  # 0.2
     return ((output.data <= eps).prod(0) > 0).any() \
-            and alignment.data[:, alignment.shape[1] - int(alignment.shape[1]/9):].sum() > 0.7
+            and alignment.data[:, -10:].sum() > 0.7