Added functionality of passing loss reduction as argument for RNNClassifier with default changed to SUM_OVER_BATCH_SIZE

This would involve making changes to all existing uses of RNNClassifier to set the loss reduction argument explicitly to SUM (previous default was SUM)

PiperOrigin-RevId: 211917502

3 files changed, 39 insertions, 17 deletions

diff --git a/tensorflow/contrib/estimator/BUILD b/tensorflow/contrib/estimator/BUILD
index 77f62df99d..437b3d965d 100644
--- a/tensorflow/contrib/estimator/BUILD
+++ b/tensorflow/contrib/estimator/BUILD
@@ -446,6 +446,7 @@ py_library(
         "//tensorflow/python/estimator",
         "//tensorflow/python/estimator:head",
         "//tensorflow/python/estimator:optimizers",
+        "//tensorflow/python/ops/losses",
         "@six_archive//:six",
     ],
 )
diff --git a/tensorflow/contrib/estimator/python/estimator/rnn.py b/tensorflow/contrib/estimator/python/estimator/rnn.py
index 7c49cd00d1..98660bb731 100644
--- a/tensorflow/contrib/estimator/python/estimator/rnn.py
+++ b/tensorflow/contrib/estimator/python/estimator/rnn.py
@@ -37,6 +37,7 @@ from tensorflow.python.ops import partitioned_variables
 from tensorflow.python.ops import rnn
 from tensorflow.python.ops import rnn_cell
 from tensorflow.python.ops import variable_scope
+from tensorflow.python.ops.losses import losses
 from tensorflow.python.summary import summary
 from tensorflow.python.training import optimizer as optimizer_lib
 from tensorflow.python.training import training_util
@@ -405,6 +406,7 @@ class RNNClassifier(estimator.Estimator):
                weight_column=None,
                label_vocabulary=None,
                optimizer='Adagrad',
+               loss_reduction=losses.Reduction.SUM_OVER_BATCH_SIZE,
                input_layer_partitioner=None,
                config=None):
     """Initializes a `RNNClassifier` instance.
@@ -454,6 +456,8 @@ class RNNClassifier(estimator.Estimator):
         string.
       optimizer: An instance of `tf.Optimizer` or string specifying optimizer
         type. Defaults to Adagrad optimizer.
+      loss_reduction: One of `tf.losses.Reduction` except `NONE`. Describes how
+        to reduce training loss over batch. Defaults to `SUM_OVER_BATCH_SIZE`.
       input_layer_partitioner: Optional. Partitioner for input layer. Defaults
         to `min_max_variable_partitioner` with `min_slice_size` 64 << 20.
       config: `RunConfig` object to configure the runtime settings.
@@ -467,11 +471,15 @@ class RNNClassifier(estimator.Estimator):
     if n_classes == 2:
       head = head_lib._binary_logistic_head_with_sigmoid_cross_entropy_loss(  # pylint: disable=protected-access
           weight_column=weight_column,
-          label_vocabulary=label_vocabulary)
+          label_vocabulary=label_vocabulary,
+          loss_reduction=loss_reduction)
     else:
       head = head_lib._multi_class_head_with_softmax_cross_entropy_loss(  # pylint: disable=protected-access
-          n_classes, weight_column=weight_column,
-          label_vocabulary=label_vocabulary)
+          n_classes,
+          weight_column=weight_column,
+          label_vocabulary=label_vocabulary,
+          loss_reduction=loss_reduction)
+
     def _model_fn(features, labels, mode, config):
       return _rnn_model_fn(
           features=features,
diff --git a/tensorflow/contrib/estimator/python/estimator/rnn_test.py b/tensorflow/contrib/estimator/python/estimator/rnn_test.py
index 959b40371a..1aebed348d 100644
--- a/tensorflow/contrib/estimator/python/estimator/rnn_test.py
+++ b/tensorflow/contrib/estimator/python/estimator/rnn_test.py
@@ -713,7 +713,7 @@ class RNNClassifierTrainingTest(test.TestCase):
 
     # Uses same checkpoint and examples as testBinaryClassEvaluationMetrics.
     # See that test for loss calculation.
-    mock_optimizer = self._mock_optimizer(expected_loss=1.119661)
+    mock_optimizer = self._mock_optimizer(expected_loss=0.559831)
 
     sequence_feature_columns = [
         seq_fc.sequence_numeric_column('price', shape=(1,))]
@@ -748,7 +748,7 @@ class RNNClassifierTrainingTest(test.TestCase):
 
     # Uses same checkpoint and examples as testMultiClassEvaluationMetrics.
     # See that test for loss calculation.
-    mock_optimizer = self._mock_optimizer(expected_loss=2.662932)
+    mock_optimizer = self._mock_optimizer(expected_loss=1.331465)
 
     sequence_feature_columns = [
         seq_fc.sequence_numeric_column('price', shape=(1,))]
@@ -812,20 +812,32 @@ class RNNClassifierEvaluationTest(test.TestCase):
     # probability = exp(logits) / (1 + exp(logits)) = [[0.353593], [0.504930]]
     # loss = -label * ln(p) - (1 - label) * ln(1 - p)
     #      = [[0.436326], [0.683335]]
+    # sum_over_batch_size = (0.436326 + 0.683335)/2
     expected_metrics = {
-        ops.GraphKeys.GLOBAL_STEP: global_step,
-        metric_keys.MetricKeys.LOSS: 1.119661,
-        metric_keys.MetricKeys.LOSS_MEAN: 0.559831,
-        metric_keys.MetricKeys.ACCURACY: 1.0,
-        metric_keys.MetricKeys.PREDICTION_MEAN: 0.429262,
-        metric_keys.MetricKeys.LABEL_MEAN: 0.5,
-        metric_keys.MetricKeys.ACCURACY_BASELINE: 0.5,
+        ops.GraphKeys.GLOBAL_STEP:
+            global_step,
+        metric_keys.MetricKeys.LOSS:
+            0.559831,
+        metric_keys.MetricKeys.LOSS_MEAN:
+            0.559831,
+        metric_keys.MetricKeys.ACCURACY:
+            1.0,
+        metric_keys.MetricKeys.PREDICTION_MEAN:
+            0.429262,
+        metric_keys.MetricKeys.LABEL_MEAN:
+            0.5,
+        metric_keys.MetricKeys.ACCURACY_BASELINE:
+            0.5,
         # With default threshold of 0.5, the model is a perfect classifier.
-        metric_keys.MetricKeys.RECALL: 1.0,
-        metric_keys.MetricKeys.PRECISION: 1.0,
+        metric_keys.MetricKeys.RECALL:
+            1.0,
+        metric_keys.MetricKeys.PRECISION:
+            1.0,
         # Positive example is scored above negative, so AUC = 1.0.
-        metric_keys.MetricKeys.AUC: 1.0,
-        metric_keys.MetricKeys.AUC_PR: 1.0,
+        metric_keys.MetricKeys.AUC:
+            1.0,
+        metric_keys.MetricKeys.AUC_PR:
+            1.0,
     }
     self.assertAllClose(
         sorted_key_dict(expected_metrics), sorted_key_dict(eval_metrics))
@@ -871,9 +883,10 @@ class RNNClassifierEvaluationTest(test.TestCase):
     #                          [0.059494, 0.572639, 0.367866]]
     # loss = -1. * log(softmax[label])
     #      = [[2.105432], [0.557500]]
+    # sum_over_batch_size = (2.105432 + 0.557500)/2
     expected_metrics = {
         ops.GraphKeys.GLOBAL_STEP: global_step,
-        metric_keys.MetricKeys.LOSS: 2.662932,
+        metric_keys.MetricKeys.LOSS: 1.331465,
         metric_keys.MetricKeys.LOSS_MEAN: 1.331466,
         metric_keys.MetricKeys.ACCURACY: 0.5,
     }