make return wave samples not contain zeros

which make training unstable

data_loader/data_loaders.py

@@ -79,7 +79,8 @@ def __init__(self,
         self.waves = []
         self.sr = None
         self.files = []
-        self.file_lengths = []
+
+        file_lengths = []
 
         def get_nframes(info_str):
             try:
@@ -96,26 +97,26 @@ def get_nframes(info_str):
                 filename = os.path.join(self.data_path, f)
                 f_obj = sf.SoundFile(filename)
                 self.files.append(f_obj)
-                self.file_lengths.append(get_nframes(f_obj.extra_info))
+                file_lengths.append(max(0, get_nframes(f_obj.extra_info) - segment) + 1)
 
                 if not self.sr:
                     self.sr = f_obj.samplerate
                 else:
                     assert f_obj.samplerate == self.sr
-        self.file_lengths = np.array(self.file_lengths)
-        self.boundaries = np.cumsum(self.file_lengths) / (self.file_lengths.sum() - 1)
+        self.file_lengths = np.array(file_lengths)
+        self.boundaries = np.cumsum(self.file_lengths) / self.file_lengths.sum()
 
         # normalization value based on each file
         # will updated on the fily
-        self.max_values = np.zeros_like(self.boundaries)
+        self.max_values = np.zeros(self.file_lengths.shape).astype(np.float32)
 
     def __len__(self):
         return self.size
 
     def __getitem__(self, index):
-        index = np.digitize(random.uniform(0, 1), self.boundaries)
+        index = np.digitize(random.uniform(0, 1), self.boundaries, right=True)
         f, length = self.files[index], self.file_lengths[index]
-        pos = random.randint(0, length - 1)
+        pos = random.randrange(0, length)
         f.seek(pos)
         x = f.read(self.segment, dtype='float32', always_2d=True, fill_value=0.).mean(1)
         max_abs = np.abs(x).max()