Merge pull request #7 from candlewill/patch-3

use clipped_gradients not the original gradients
keithito · Jul 24, 2017 · aca83c8 · aca83c8
2 parents cdf389b + 9951817
commit aca83c8
Showing 1 changed file with 1 addition and 1 deletion.
diff --git a/models/tacotron.py b/models/tacotron.py
@@ -138,7 +138,7 @@ def add_optimizer(self, global_step):
       # Add dependency on UPDATE_OPS; otherwise batchnorm won't work correctly. See:
       # https://github.com/tensorflow/tensorflow/issues/1122
       with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
-        self.optimize = optimizer.apply_gradients(zip(gradients, variables),
+        self.optimize = optimizer.apply_gradients(zip(clipped_gradients, variables),
           global_step=global_step)