Change public fns to support non-tensor inputs.

Fix `index_table_from_tensor` to handle `int32` keys with no oov buckets.
Add better tests for sparse integer inputs.
Change: 150006782

tensorflow/contrib/layers/python/layers/feature_column_ops.py

@@ -136,6 +136,19 @@ def _embeddings_from_arguments(column,
       max_norm=args.max_norm)
 
 
+def _convert_to_tensors(values):
+  """Modifies `values`, if necessary, to convert all to `Tensor` objects.
+
+  Args:
+    values: Dict mapping a key to tensor-like values.
+  """
+  for k in sorted(
+      # pylint: disable=protected-access
+      values.keys(),
+      key=lambda k: k.key if isinstance(k, fc._FeatureColumn) else k):
+    values[k] = sparse_tensor_py.convert_to_tensor_or_sparse_tensor(values[k])
+
+
 def _input_from_feature_columns(columns_to_tensors,
                                 feature_columns,
                                 weight_collections,
@@ -148,6 +161,7 @@ def _input_from_feature_columns(columns_to_tensors,
   with variable_scope.variable_scope(scope,
                                      default_name=default_name,
                                      values=columns_to_tensors.values()):
+    _convert_to_tensors(columns_to_tensors)
     output_tensors = []
     transformer = _Transformer(columns_to_tensors)
     if weight_collections:
@@ -226,7 +240,8 @@ def input_from_feature_columns(columns_to_tensors,
     columns_to_tensors: A mapping from feature column to tensors. 'string' key
       means a base feature (not-transformed). It can have FeatureColumn as a
       key too. That means that FeatureColumn is already transformed by input
-      pipeline. For example, `inflow` may have handled transformations.
+      pipeline. For example, `inflow` may have handled transformations. This
+      dictionary will be modified by this function.
     feature_columns: A set containing all the feature columns. All items in the
       set should be instances of classes derived by FeatureColumn.
     weight_collections: List of graph collections to which weights are added.
@@ -235,7 +250,7 @@ def input_from_feature_columns(columns_to_tensors,
     scope: Optional scope for variable_scope.
 
   Returns:
-    A Tensor which can be consumed by hidden layers in the neural network.
+    `Tensor` which can be consumed by hidden layers in the neural network.
 
   Raises:
     ValueError: if FeatureColumn cannot be consumed by a neural network.
@@ -268,7 +283,8 @@ def sequence_input_from_feature_columns(columns_to_tensors,
     columns_to_tensors: A mapping from feature column to tensors. 'string' key
       means a base feature (not-transformed). It can have FeatureColumn as a
       key too. That means that FeatureColumn is already transformed by input
-      pipeline. For example, `inflow` may have handled transformations.
+      pipeline. For example, `inflow` may have handled transformations. This
+      dictionary will be modified by this function.
     feature_columns: A set containing all the feature columns. All items in the
       set should be instances of classes derived by FeatureColumn.
     weight_collections: List of graph collections to which weights are added.
@@ -410,7 +426,8 @@ def joint_weighted_sum_from_feature_columns(columns_to_tensors,
     columns_to_tensors: A mapping from feature column to tensors. 'string' key
       means a base feature (not-transformed). It can have FeatureColumn as a
       key too. That means that FeatureColumn is already transformed by input
-      pipeline. For example, `inflow` may have handled transformations.
+      pipeline. For example, `inflow` may have handled transformations. This
+      dictionary will be modified by this function.
     feature_columns: A set containing all the feature columns. All items in the
       set should be instances of classes derived from FeatureColumn.
     num_outputs: An integer specifying number of outputs. Default value is 1.
@@ -435,6 +452,7 @@ def joint_weighted_sum_from_feature_columns(columns_to_tensors,
       scope,
       default_name='joint_weighted_sum_from_feature_columns',
       values=columns_to_tensors.values()):
+    _convert_to_tensors(columns_to_tensors)
     transformer = _Transformer(columns_to_tensors)
     embedding_lookup_arguments = []
     for column in sorted(set(feature_columns), key=lambda x: x.key):
@@ -499,7 +517,8 @@ def weighted_sum_from_feature_columns(columns_to_tensors,
     columns_to_tensors: A mapping from feature column to tensors. 'string' key
       means a base feature (not-transformed). It can have FeatureColumn as a
       key too. That means that FeatureColumn is already transformed by input
-      pipeline. For example, `inflow` may have handled transformations.
+      pipeline. For example, `inflow` may have handled transformations. This
+      dictionary will be modified by this function.
     feature_columns: A set containing all the feature columns. All items in the
       set should be instances of classes derived from FeatureColumn.
     num_outputs: An integer specifying number of outputs. Default value is 1.
@@ -523,6 +542,7 @@ def weighted_sum_from_feature_columns(columns_to_tensors,
       scope,
       default_name='weighted_sum_from_feature_columns',
       values=columns_to_tensors.values()):
+    _convert_to_tensors(columns_to_tensors)
     output_tensors = []
     column_to_variable = dict()
     transformer = _Transformer(columns_to_tensors)
@@ -677,7 +697,8 @@ def transform_features(features, feature_columns):
   ```
 
   Args:
-    features: A dictionary of features.
+    features: A dictionary mapping string names to `Tensor` or `SparseTensor`
+      feature values. This dictionary will not be modified by this function.
     feature_columns: An iterable containing all the feature columns. All items
       should be instances of classes derived from _FeatureColumn.
 
@@ -685,15 +706,15 @@ def transform_features(features, feature_columns):
     A `dict` mapping FeatureColumn to `Tensor` and `SparseTensor` values.
   """
   check_feature_columns(feature_columns)
-  columns_to_tensor = features.copy()
-  transformer = _Transformer(columns_to_tensor)
+  columns_to_tensors = features.copy()
+  _convert_to_tensors(columns_to_tensors)
+  transformer = _Transformer(columns_to_tensors)
   for column in sorted(set(feature_columns), key=lambda x: x.key):
     transformer.transform(column)
-  keys = list(columns_to_tensor.keys())
-  for k in keys:
+  for k in list(columns_to_tensors.keys()):
     if k not in feature_columns:
-      columns_to_tensor.pop(k)
-  return columns_to_tensor
+      columns_to_tensors.pop(k)
+  return columns_to_tensors
 
 
 def parse_feature_columns_from_sequence_examples(
@@ -765,6 +786,7 @@ def _log_variable(variable):
 
 def _infer_real_valued_column_for_tensor(name, tensor):
   """Creates a real_valued_column for given tensor and name."""
+  tensor = sparse_tensor_py.convert_to_tensor_or_sparse_tensor(tensor)
   if isinstance(tensor, sparse_tensor_py.SparseTensor):
     raise ValueError(
         'SparseTensor is not supported for auto detection. Please define '
@@ -817,6 +839,13 @@ def check_feature_columns(feature_columns):
     seen_keys.add(key)
 
 
+def _sort_columns(columns_or_strings):
+  return sorted(
+      # pylint: disable=protected-access
+      columns_or_strings,
+      key=lambda k: k.key if isinstance(k, fc._FeatureColumn) else k)
+
+
 class _Transformer(object):
   """Handles all the transformations defined by FeatureColumn if needed.
 
@@ -874,7 +903,7 @@ class _Transformer(object):
     Raises:
       ValueError: if FeatureColumn cannot be handled by this Transformer.
     """
-    logging.debug('Transforming feature_column %s', feature_column)
+    logging.debug('Transforming feature_column %s.', feature_column)
     if feature_column in self._columns_to_tensors:
       # Feature_column is already transformed.
       return self._columns_to_tensors[feature_column]

tensorflow/contrib/lookup/lookup_ops.py

@@ -1029,13 +1029,12 @@ def index_table_from_tensor(mapping,
           name="table_init")
       table = HashTable(
           init, default_value, shared_name=shared_name, name=hash_table_scope)
-    if num_oov_buckets:
-      table = IdTableWithHashBuckets(
-          table,
-          num_oov_buckets=num_oov_buckets,
-          hasher_spec=hasher_spec,
-          name=feat_to_id_scope,
-          key_dtype=dtype)
+    table = IdTableWithHashBuckets(
+        table,
+        num_oov_buckets=num_oov_buckets,
+        hasher_spec=hasher_spec,
+        name=feat_to_id_scope,
+        key_dtype=dtype)
 
     return table
 

tensorflow/contrib/lookup/lookup_ops_test.py

@@ -1341,6 +1341,17 @@ class IndexTableFromTensor(test.TestCase):
       data_flow_ops.tables_initializer().run()
       self.assertAllEqual((1, 2, 3), ids.eval())
 
+  def test_int32_index_table_from_tensor_with_no_buckets(self):
+    with self.test_session():
+      table = lookup.index_table_from_tensor(
+          mapping=(42, 1, -1000), dtype=dtypes.int32)
+      ids = table.lookup(
+          constant_op.constant((1, -1000, 11), dtype=dtypes.int32))
+
+      self.assertRaises(errors_impl.OpError, ids.eval)
+      data_flow_ops.tables_initializer().run()
+      self.assertAllEqual((1, 2, -1), ids.eval())
+
   def test_int64_index_table_from_tensor_with_tensor_init(self):
     with self.test_session():
       table = lookup.index_table_from_tensor(