Merge pull request tensorflow#15632 from drpngx/branch_180053468

caisq · web-flow · commit b95724bf529c · 2017-12-25T23:53:51.000-05:00
Branch 180053468
diff --git a/tensorflow/contrib/cudnn_rnn/python/kernel_tests/cudnn_rnn_test.py b/tensorflow/contrib/cudnn_rnn/python/kernel_tests/cudnn_rnn_test.py
@@ -314,6 +314,47 @@ def testLayerBasic(self):
       self.assertEqual(0, total_sum2_v)
       self.assertEqual(0, total_sum3_v)
 
+  def testSaveableGraphDeviceAssignment(self):
+    num_layers = 4
+    num_units = 2
+    batch_size = 8
+    direction = CUDNN_RNN_UNIDIRECTION
+    dir_count = 1
+
+    def DeviceFn(op):
+      if op.type in ("Variable", "VariableV2"):
+        return "/cpu:0"
+      else:
+        return "/gpu:0"
+
+    with ops.Graph().as_default() as g:
+      with ops.device(DeviceFn):
+        with vs.variable_scope("main"):
+          kernel_initializer = init_ops.constant_initializer(3.14)
+          bias_initializer = init_ops.constant_initializer(1.59)
+          inputs = random_ops.random_uniform(
+              [num_layers * dir_count, batch_size, num_units],
+              dtype=dtypes.float32)
+
+          lstm = cudnn_rnn.CudnnLSTM(num_layers, num_units,
+                                     direction=direction,
+                                     kernel_initializer=kernel_initializer,
+                                     bias_initializer=bias_initializer,
+                                     name="awesome_lstm")
+          outputs = lstm(inputs)
+
+        # saver is created in the scope of DeviceFn.
+        saver = saver_lib.Saver()
+
+    with self.test_session(use_gpu=True, graph=g) as sess:
+      save_path = os.path.join(self.get_temp_dir(),
+                               "test-saveable-device-assignment")
+      sess.run(variables.global_variables_initializer())
+
+      saver.save(sess, save_path)
+      saver.restore(sess, save_path)
+      sess.run(outputs)
+
 
 # TODO(jamesqin): Transform to parameterized test after it is included in the
 # TF open source codebase.
diff --git a/tensorflow/contrib/cudnn_rnn/python/layers/cudnn_rnn.py b/tensorflow/contrib/cudnn_rnn/python/layers/cudnn_rnn.py
@@ -450,17 +450,18 @@ def _canonical_to_opaque(self, cu_weights, cu_biases):
       raise RuntimeError(
           "%s._canonical_to_opaque invoked before input shape is known" %
           type(self).__name__)
-    return cudnn_rnn_ops.cudnn_rnn_canonical_to_opaque_params(
-        rnn_mode=self._rnn_mode,
-        num_layers=self._num_layers,
-        num_units=self._num_units,
-        input_size=self._input_size,
-        weights=cu_weights,
-        biases=cu_biases,
-        input_mode=self._input_mode,
-        seed=self._seed,
-        dropout=self._dropout,
-        direction=self._direction)
+    with ops.device("/gpu:0"):
+      return cudnn_rnn_ops.cudnn_rnn_canonical_to_opaque_params(
+          rnn_mode=self._rnn_mode,
+          num_layers=self._num_layers,
+          num_units=self._num_units,
+          input_size=self._input_size,
+          weights=cu_weights,
+          biases=cu_biases,
+          input_mode=self._input_mode,
+          seed=self._seed,
+          dropout=self._dropout,
+          direction=self._direction)
 
   def _forward(self, inputs, h, c, opaque_params, training):
     output, output_h, output_c = cudnn_rnn_ops._cudnn_rnn(  # pylint:disable=protected-access
@@ -489,12 +490,12 @@ def _create_saveable(self):
     if self._saveable is not None:
       raise RuntimeError("Cudnn saveable already created.")
     self._saveable = self._saveable_cls(  # pylint:disable=not-callable
-        self.trainable_variables[0],
-        self.num_layers,
-        self.num_units,
-        self.input_size,
-        self.input_mode,
-        self.direction,
+        opaque_params=self.trainable_variables[0],
+        num_layers=self.num_layers,
+        num_units=self.num_units,
+        input_size=self.input_size,
+        input_mode=self.input_mode,
+        direction=self.direction,
         scope=vs.get_variable_scope(),
         name="%s_saveable" % self.trainable_variables[0].op.name)
     ops.add_to_collection(ops.GraphKeys.SAVEABLE_OBJECTS, self._saveable)
diff --git a/tensorflow/contrib/cudnn_rnn/python/ops/cudnn_rnn_ops.py b/tensorflow/contrib/cudnn_rnn/python/ops/cudnn_rnn_ops.py
@@ -72,7 +72,7 @@ class CudnnCompatibleLSTMCell(lstm_ops.LSTMBlockCell):
   def __init__(self, num_units, reuse=None):
     super(CudnnCompatibleLSTMCell, self).__init__(
         num_units, forget_bias=0, cell_clip=None, use_peephole=False,
-        reuse=reuse)
+        reuse=reuse, name="cudnn_compatible_lstm_cell")
     self._names.update({"scope": "cudnn_compatible_lstm_cell"})
 
 
@@ -303,16 +303,17 @@ def _OpaqueParamsToCanonical(self):
     Returns:
       2 list for weights and biases respectively.
     """
-    weights, biases = gen_cudnn_rnn_ops.cudnn_rnn_params_to_canonical(
-        num_layers=self._num_layers,
-        num_units=self._num_units,
-        input_size=self._input_size,
-        params=self._variables,
-        num_params=self._num_params,
-        rnn_mode=self._rnn_mode,
-        input_mode=self._input_mode,
-        direction=self._direction)
-    return (weights, biases)
+    with ops.device("/gpu:0"):
+      weights, biases = gen_cudnn_rnn_ops.cudnn_rnn_params_to_canonical(
+          num_layers=self._num_layers,
+          num_units=self._num_units,
+          input_size=self._input_size,
+          params=self._variables,
+          num_params=self._num_params,
+          rnn_mode=self._rnn_mode,
+          input_mode=self._input_mode,
+          direction=self._direction)
+      return (weights, biases)
 
   def _CanonicalToOpaqueParams(self, cu_weights, cu_biases):
     """Converts from Cudnn canonical format to opaque params.
@@ -323,15 +324,16 @@ def _CanonicalToOpaqueParams(self, cu_weights, cu_biases):
     Returns:
       a single opaque tensor.
     """
-    return gen_cudnn_rnn_ops.cudnn_rnn_canonical_to_params(
-        num_layers=self._num_layers,
-        num_units=self._num_units,
-        input_size=self._input_size,
-        weights=cu_weights,
-        biases=cu_biases,
-        rnn_mode=self._rnn_mode,
-        input_mode=self._input_mode,
-        direction=self._direction)
+    with ops.device("/gpu:0"):
+      return gen_cudnn_rnn_ops.cudnn_rnn_canonical_to_params(
+          num_layers=self._num_layers,
+          num_units=self._num_units,
+          input_size=self._input_size,
+          weights=cu_weights,
+          biases=cu_biases,
+          rnn_mode=self._rnn_mode,
+          input_mode=self._input_mode,
+          direction=self._direction)
 
   def _TransformCanonical(self, cu_weights, cu_biases):
     r"""Transform from Cudnn canonical to tf canonical.
