Fix the colocate_with issue for Adagrad optimizerV2.

PiperOrigin-RevId: 212702577

3 files changed, 22 insertions, 11 deletions

diff --git a/tensorflow/contrib/distribute/python/combinations.py b/tensorflow/contrib/distribute/python/combinations.py
index 1133be6d0b..244d1fcec8 100644
--- a/tensorflow/contrib/distribute/python/combinations.py
+++ b/tensorflow/contrib/distribute/python/combinations.py
@@ -50,10 +50,12 @@ from tensorflow.contrib.cluster_resolver import TPUClusterResolver
 from tensorflow.contrib.distribute.python import mirrored_strategy as mirrored_lib
 from tensorflow.contrib.distribute.python import one_device_strategy as one_device_lib
 from tensorflow.contrib.distribute.python import tpu_strategy as tpu_lib
+from tensorflow.contrib.optimizer_v2 import adagrad as adagrad_v2
 from tensorflow.contrib.optimizer_v2 import adam as adam_v2
 from tensorflow.contrib.optimizer_v2 import gradient_descent as gradient_descent_v2
 from tensorflow.python.eager import context
 from tensorflow.python.framework import ops
+from tensorflow.python.training import adagrad
 from tensorflow.python.training import adam
 from tensorflow.python.training import distribution_strategy_context
 from tensorflow.python.training import gradient_descent
@@ -347,17 +349,23 @@ mirrored_strategy_with_two_gpus = NamedDistribution(
 
 
 adam_optimizer_v1_fn = NamedObject(
-    "AdamV1", lambda: adam.AdamOptimizer(0.2, epsilon=1))
+    "AdamV1", lambda: adam.AdamOptimizer(0.001, epsilon=1))
 gradient_descent_optimizer_v1_fn = NamedObject(
     "GradientDescentV1", lambda: gradient_descent.GradientDescentOptimizer(0.2))
-optimizers_v1 = [adam_optimizer_v1_fn, gradient_descent_optimizer_v1_fn]
+adagrad_optimizer_v1_fn = NamedObject(
+    "AdagradV1", lambda: adagrad.AdagradOptimizer(0.001))
+optimizers_v1 = [adam_optimizer_v1_fn, gradient_descent_optimizer_v1_fn,
+                 adagrad_optimizer_v1_fn]
 
 adam_optimizer_v2_fn = NamedObject(
-    "AdamV2", lambda: adam_v2.AdamOptimizer(0.2, epsilon=1))
+    "AdamV2", lambda: adam_v2.AdamOptimizer(0.001, epsilon=1))
 gradient_descent_optimizer_v2_fn = NamedObject(
     "GradientDescentV2",
     lambda: gradient_descent_v2.GradientDescentOptimizer(0.2))
-optimizers_v2 = [adam_optimizer_v2_fn, gradient_descent_optimizer_v2_fn]
+adagrad_optimizer_v2_fn = NamedObject(
+    "AdagradV2", lambda: adagrad_v2.AdagradOptimizer(0.001))
+optimizers_v2 = [adam_optimizer_v2_fn, gradient_descent_optimizer_v2_fn,
+                 adagrad_optimizer_v2_fn]
 
 graph_and_eager_modes = ["graph", "eager"]
 
diff --git a/tensorflow/contrib/distribute/python/minimize_loss_test.py b/tensorflow/contrib/distribute/python/minimize_loss_test.py
index bdac4fb58c..ba147e7824 100644
--- a/tensorflow/contrib/distribute/python/minimize_loss_test.py
+++ b/tensorflow/contrib/distribute/python/minimize_loss_test.py
@@ -183,6 +183,10 @@ class MinimizeLossStepTest(test.TestCase, parameterized.TestCase):
                 "dense/kernel", "dense/bias", "beta1_power", "beta2_power",
                 "dense/kernel/Adam", "dense/kernel/Adam_1", "dense/bias/Adam",
                 "dense/bias/Adam_1"
+            ],
+            "Adagrad": [
+                "dense/kernel/Adagrad", "dense/kernel",
+                "dense/bias/Adagrad", "dense/bias"
             ]
         }
         variables = variables_map[optimizer_fn().get_name()]
diff --git a/tensorflow/contrib/optimizer_v2/adagrad.py b/tensorflow/contrib/optimizer_v2/adagrad.py
index c333d1e089..25ec475499 100644
--- a/tensorflow/contrib/optimizer_v2/adagrad.py
+++ b/tensorflow/contrib/optimizer_v2/adagrad.py
@@ -64,18 +64,17 @@ class AdagradOptimizer(optimizer_v2.OptimizerV2):
 
   def _create_vars(self, var_list, state):
     for v in var_list:
-      # TODO(isaprykin): Delete colocate_with(v) from other optimizers and
-      # confirm that colocation will happen anyway.
       dtype = v.dtype.base_dtype
       if v.get_shape().is_fully_defined():
         init = init_ops.constant_initializer(self._initial_accumulator_value,
                                              dtype=dtype)
       else:
-        # Use a Tensor instead of initializer if variable does not have static
-        # shape.
-        init_constant = gen_array_ops.fill(
-            array_ops.shape(v), self._initial_accumulator_value)
-        init = math_ops.cast(init_constant, dtype)
+        def init(v=v, dtype=dtype):
+          # Use a Tensor instead of initializer if variable does not have
+          # static shape.
+          init_constant = gen_array_ops.fill(array_ops.shape(v),
+                                             self._initial_accumulator_value)
+          return math_ops.cast(init_constant, dtype)
       state.create_slot_with_initializer(v, init, v.get_shape(), dtype,
                                          "accumulator")