Data loader bug fix 2

datasets/LJSpeech.py

@@ -125,6 +125,7 @@ class LJSpeechDataset(Dataset):
             mel = torch.FloatTensor(mel)
             mel_lengths = torch.LongTensor(mel_lengths)
             stop_targets = torch.FloatTensor(stop_targets)
+            
             return text, text_lenghts, linear, mel, mel_lengths, stop_targets, item_idxs[0]
 
         raise TypeError(("batch must contain tensors, numbers, dicts or lists;\

layers/tacotron.py

@@ -256,9 +256,12 @@ class Decoder(nn.Module):
         greedy = not self.training
 
         if memory is not None:
+            print(memory.shape)
+            
             # Grouping multiple frames if necessary
             if memory.size(-1) == self.memory_dim:
                 memory = memory.view(B, memory.size(1) // self.r, -1)
+            print(memory.shape)
             assert memory.size(-1) == self.memory_dim * self.r,\
                 " !! Dimension mismatch {} vs {} * {}".format(memory.size(-1),
                                                          self.memory_dim, self.r)

train.py

@@ -82,7 +82,7 @@ def train(model, criterion, data_loader, optimizer, epoch):
         linear_input = data[2]
         mel_input = data[3]
         mel_lengths = data[4]
-
+        
         current_step = num_iter + args.restore_step + epoch * len(data_loader) + 1
 
         # setup lr

utils/data.py

@@ -17,13 +17,13 @@ def prepare_data(inputs):
 def _pad_tensor(x, length):
     _pad = 0
     assert x.ndim == 2
-    return np.pad(x, [[0, 0], [0, length - x.shape[1]]], mode='constant', constant_values=_pad)
-
+    x = np.pad(x, [[0, 0], [0, length - x.shape[1]]], mode='constant', constant_values=_pad)
+    return x
 
 def prepare_tensor(inputs, out_steps):
     max_len = max((x.shape[1] for x in inputs)) + 1 # zero-frame
     remainder = max_len % out_steps
-    return np.stack([_pad_tensor(x, max_len + remainder) for x in inputs])
+    return np.stack([_pad_tensor(x, max_len + (out_steps - remainder)) for x in inputs])
 
 
 def _pad_stop_target(x, length):
@@ -35,7 +35,7 @@ def _pad_stop_target(x, length):
 def prepare_stop_target(inputs, out_steps):
     max_len = max((x.shape[0] for x in inputs)) + 1  # zero-frame
     remainder = max_len % out_steps
-    return np.stack([_pad_stop_target(x, max_len + remainder) for x in inputs])
+    return np.stack([_pad_stop_target(x, max_len + (out_steps - remainder)) for x in inputs])
 
 
 def pad_per_step(inputs, pad_len):