From e73199d413395d7ede0dd86c06f7a9ad2d1fee17 Mon Sep 17 00:00:00 2001
From: Gabriel de Marmiesse <gabrieldemarmiesse@gmail.com>
Date: Wed, 11 Apr 2018 07:21:51 +0200
Subject: [PATCH] Removed generate dropout ones from recurrent. (#9892)

* Removed generate dropout ones from recurrent.

* Fixed index issue.
---
 keras/layers/recurrent.py | 22 ++++++----------------
 1 file changed, 6 insertions(+), 16 deletions(-)

diff --git a/keras/layers/recurrent.py b/keras/layers/recurrent.py
index 952c6c77ce8..cbe6ce9a7e3 100644
--- a/keras/layers/recurrent.py
+++ b/keras/layers/recurrent.py
@@ -888,13 +888,13 @@ def call(self, inputs, states, training=None):
         prev_output = states[0]
         if 0 < self.dropout < 1 and self._dropout_mask is None:
             self._dropout_mask = _generate_dropout_mask(
-                _generate_dropout_ones(inputs, K.shape(inputs)[-1]),
+                K.ones_like(inputs),
                 self.dropout,
                 training=training)
         if (0 < self.recurrent_dropout < 1 and
                 self._recurrent_dropout_mask is None):
             self._recurrent_dropout_mask = _generate_dropout_mask(
-                _generate_dropout_ones(inputs, self.units),
+                K.ones_like(prev_output),
                 self.recurrent_dropout,
                 training=training)
 
@@ -1329,14 +1329,14 @@ def call(self, inputs, states, training=None):
 
         if 0 < self.dropout < 1 and self._dropout_mask is None:
             self._dropout_mask = _generate_dropout_mask(
-                _generate_dropout_ones(inputs, K.shape(inputs)[-1]),
+                K.ones_like(inputs),
                 self.dropout,
                 training=training,
                 count=3)
         if (0 < self.recurrent_dropout < 1 and
                 self._recurrent_dropout_mask is None):
             self._recurrent_dropout_mask = _generate_dropout_mask(
-                _generate_dropout_ones(inputs, self.units),
+                K.ones_like(h_tm1),
                 self.recurrent_dropout,
                 training=training,
                 count=3)
@@ -1887,14 +1887,14 @@ def bias_initializer(_, *args, **kwargs):
     def call(self, inputs, states, training=None):
         if 0 < self.dropout < 1 and self._dropout_mask is None:
             self._dropout_mask = _generate_dropout_mask(
-                _generate_dropout_ones(inputs, K.shape(inputs)[-1]),
+                K.ones_like(inputs),
                 self.dropout,
                 training=training,
                 count=4)
         if (0 < self.recurrent_dropout < 1 and
                 self._recurrent_dropout_mask is None):
             self._recurrent_dropout_mask = _generate_dropout_mask(
-                _generate_dropout_ones(inputs, self.units),
+                K.ones_like(states[0]),
                 self.recurrent_dropout,
                 training=training,
                 count=4)
@@ -2249,16 +2249,6 @@ def from_config(cls, config):
         return cls(**config)
 
 
-def _generate_dropout_ones(inputs, dims):
-    # Currently, CNTK can't instantiate `ones` with symbolic shapes.
-    # Will update workaround once CNTK supports it.
-    if K.backend() == 'cntk':
-        ones = K.ones_like(K.reshape(inputs[:, 0], (-1, 1)))
-        return K.tile(ones, (1, dims))
-    else:
-        return K.ones((K.shape(inputs)[0], dims))
-
-
 def _generate_dropout_mask(ones, rate, training=None, count=1):
     def dropped_inputs():
         return K.dropout(ones, rate)