Change to in FastSpeechPostnet, may cause nan problem.

dathudeptrai · dathudeptrai · commit 4cbdd05482a6 · 2020-08-06T10:24:55.000+07:00
diff --git a/tensorflow_tts/models/fastspeech.py b/tensorflow_tts/models/fastspeech.py
@@ -517,7 +517,7 @@ def __init__(self, config, **kwargs):
                 padding="same",
                 name="conv_._{}".format(i),
             )
-            batch_norm = tf.keras.layers.BatchNormalization(
+            batch_norm = tf.keras.layers.experimental.SyncBatchNormalization(
                 name="batch_norm_._{}".format(i)
             )
             self.conv_batch_norm.append((conv, batch_norm))

Original file line number	Diff line number	Diff line change
`@@ -517,7 +517,7 @@ def __init__(self, config, **kwargs):`
`517`	`517`	`padding="same",`
`518`	`518`	`name="conv_._{}".format(i),`
`519`	`519`	`)`
`520`		`- batch_norm = tf.keras.layers.BatchNormalization(`
	`520`	`+ batch_norm = tf.keras.layers.experimental.SyncBatchNormalization(`
`521`	`521`	`name="batch_norm_._{}".format(i)`
`522`	`522`	`)`
`523`	`523`	`self.conv_batch_norm.append((conv, batch_norm))`