Merge pull request #307 from lispc/develop

fix a deep speech 2 speed bug.

deep_speech_2/data_utils/data.py

@@ -327,7 +327,7 @@ class DataGenerator(object):
         shift_len = self._rng.randint(0, batch_size - 1)
         batch_manifest = zip(*[iter(manifest[shift_len:])] * batch_size)
         self._rng.shuffle(batch_manifest)
-        batch_manifest = list(sum(batch_manifest, ()))
+        batch_manifest = [item for batch in batch_manifest for item in batch]
         if not clipped:
             res_len = len(manifest) - shift_len - len(batch_manifest)
             batch_manifest.extend(manifest[-res_len:])