Upload eduport_tts_mal.py with huggingface_hub

eduport_tts_mal.py

@@ -51,10 +51,10 @@ class SpeechDataset(Dataset):
         input_values = self.processor(waveform, sampling_rate=self.target_sampling_rate, return_tensors="pt").input_values.squeeze(0)
         
         # Pad or truncate the audio to ensure fixed length (the longest audio length)
-        if input_values.size(0) < self.max_audio_length:
-            padding_length = self.max_audio_length - input_values.size(0)
+        if input_values.size(1) < self.max_audio_length:
+            padding_length = self.max_audio_length - input_values.size(1)
             # Create a zero tensor with the same batch size (1) and the padding length along dimension 1
-            padding = torch.zeros(1, padding_length)
+            padding = torch.zeros_like(input_values[:, :padding_length])
             input_values = torch.cat([input_values, padding], dim=1)
         else:
             input_values = input_values[:, :self.max_audio_length]  # Truncate to max_audio_length
@@ -289,4 +289,4 @@ def train_model(num_epochs=10, accumulation_steps=16):
             }, 'metadata.pth')
 
 # Run the training
-train_model()
+train_model()