Add ragged tensor input support to Keras via the 'ragged' arg in keras.Input.

PiperOrigin-RevId: 254256357

Author: tensorflower-gardener

tensorflow/python/keras/BUILD

@@ -1168,6 +1168,7 @@ tf_py_test(
         "//tensorflow/python:sparse_ops",
         "//tensorflow/python:sparse_tensor",
     ],
+    shard_count = 4,
 )
 
 tf_py_test(

tensorflow/python/keras/backend.py

@@ -38,7 +38,6 @@
 from tensorflow.python.distribute import distribution_strategy_context
 from tensorflow.python.distribute import multi_worker_util
 from tensorflow.python.eager import context
-from tensorflow.python.framework import composite_tensor_utils
 from tensorflow.python.eager import function as eager_function
 from tensorflow.python.eager import lift_to_graph
 from tensorflow.python.framework import composite_tensor
@@ -70,6 +69,7 @@
 from tensorflow.python.ops import tensor_array_grad  # pylint: disable=unused-import
 from tensorflow.python.ops import tensor_array_ops
 from tensorflow.python.ops import variables as variables_module
+from tensorflow.python.ops.ragged import ragged_factory_ops
 from tensorflow.python.platform import tf_logging as logging
 from tensorflow.python.util import nest
 from tensorflow.python.util import tf_contextlib
@@ -958,7 +958,12 @@ def is_keras_tensor(x):
 
 
 @keras_export('keras.backend.placeholder')
-def placeholder(shape=None, ndim=None, dtype=None, sparse=False, name=None):
+def placeholder(shape=None,
+                ndim=None,
+                dtype=None,
+                sparse=False,
+                name=None,
+                ragged=False):
   """Instantiates a placeholder tensor and returns it.
 
   Arguments:
@@ -970,9 +975,14 @@ def placeholder(shape=None, ndim=None, dtype=None, sparse=False, name=None):
       dtype: Placeholder type.
       sparse: Boolean, whether the placeholder should have a sparse type.
       name: Optional name string for the placeholder.
+      ragged: Boolean, whether the placeholder should have a ragged type.
+          In this case, values of 'None' in the 'shape' argument represent
+          ragged dimensions. For more information about RaggedTensors, see this
+          [guide](https://www.tensorflow.org/guide/ragged_tensors).
 
   Raises:
-      ValueError: If called with eager execution.
+      ValueError: If called with eager execution
+      ValueError: If called with sparse = True and ragged = True.
 
   Returns:
       Tensor instance (with Keras metadata included).
@@ -985,6 +995,11 @@ def placeholder(shape=None, ndim=None, dtype=None, sparse=False, name=None):
       <tf.Tensor 'Placeholder_4:0' shape=(2, 4, 5) dtype=float32>
   ```
   """
+  if sparse and ragged:
+    raise ValueError(
+        'Cannot set both sparse and ragged to True when creating a placeholder.'
+    )
+
   if dtype is None:
     dtype = floatx()
   if not shape:
@@ -993,6 +1008,20 @@ def placeholder(shape=None, ndim=None, dtype=None, sparse=False, name=None):
   with get_graph().as_default():
     if sparse:
       x = array_ops.sparse_placeholder(dtype, shape=shape, name=name)
+    elif ragged:
+      ragged_rank = 0
+      for i in range(1, len(shape)):
+        if shape[i] is None:
+          ragged_rank += 1
+        else:
+          break
+      value_shape = shape[(ragged_rank + 1):]
+
+      x = ragged_factory_ops.placeholder(
+          dtype=dtype,
+          ragged_rank=ragged_rank,
+          value_shape=value_shape,
+          name=name)
     else:
       x = array_ops.placeholder(dtype, shape=shape, name=name)
   return x
@@ -1008,7 +1037,11 @@ def is_placeholder(x):
       Boolean.
   """
   try:
-    return x.op.type == 'Placeholder'
+    if isinstance(x, composite_tensor.CompositeTensor):
+      flat_components = nest.flatten(x, expand_composites=True)
+      return py_any(is_placeholder(c) for c in flat_components)
+    else:
+      return x.op.type == 'Placeholder'
   except AttributeError:
     return False
 
@@ -3108,63 +3141,6 @@ def print_tensor(x, message=''):
     logging_ops.print_v2(message, x, output_stream=sys.stdout)
     return x
 
-
-def is_tensor_or_composite_tensor(value):
-  """Test if a passed value object is a tensor-like or composite tensor."""
-  return (tensor_util.is_tensor(value) or isinstance(value, np.ndarray) or
-          composite_tensor_utils.is_composite_or_composite_value(value))
-
-
-def _try_process_scipy_sparse_input(value):
-  """Converts 'value' to a SparseTensor if it is a scipy sparse matrix.
-
-  Arguments:
-    value: An object that may have the attributes of a scipy sparse matrix.
-
-  Returns:
-    Either a SparseTensor based off of 'value' or 'value' itself.
-  """
-  try:
-    sparse_coo = value.tocoo()
-    row, col = sparse_coo.row, sparse_coo.col
-    data, shape = sparse_coo.data, sparse_coo.shape
-  except AttributeError:
-    # If we can't convert this object, it could be either a single data
-    # element (ie, a bool/int/float) which is OK to pass on, or something
-    # that we don't understand (which may or may not be OK). In either
-    # case, don't die here: the data standardization code will catch
-    # those issues.
-    return value
-
-  indices = np.concatenate((np.expand_dims(row, 1), np.expand_dims(col, 1)), 1)
-  return sparse_tensor.SparseTensor(indices, data, shape)
-
-
-def try_convert_scipy_to_sparse(values):
-  """Converts scipy sparse matrices in 'values' to SparseTensors, if possible.
-
-  Arguments:
-    values: An input or list of inputs to convert. These may be TensorLikes,
-      ndarrays, composite tensors, or scipy sparse values.
-
-  Returns:
-    An input or list of inputs where scipy sparse tensors have been converted
-    to tf.SparseTensors.
-
-  Raises:
-    ValueError: If input cannot be converted to a SparseTensor.
-  """
-  # Convert scipy sparse data into sparse tensors.
-  value_structure = values
-  values = nest.flatten(values)
-  for idx, value in enumerate(values):
-    if not is_tensor_or_composite_tensor(value):
-      values[idx] = _try_process_scipy_sparse_input(value)
-  values = nest.pack_sequence_as(value_structure, values)
-
-  return values
-
-
 # GRAPH MANIPULATION
 
 
@@ -3194,6 +3170,7 @@ def __init__(self, inputs, outputs, updates=None, name=None,
     if not isinstance(updates, (list, tuple)):
       raise TypeError('`updates` in a Keras backend function '
                       'should be a list or tuple.')
+
     self._inputs_structure = inputs
     self.inputs = nest.flatten(inputs, expand_composites=True)
     self._outputs_structure = outputs
@@ -3311,10 +3288,6 @@ def _eval_if_composite(self, tensor):
       return tensor
 
   def __call__(self, inputs):
-    inputs = try_convert_scipy_to_sparse(inputs)
-
-    # Ensure that input value types match any expected composite tensor types.
-    # TODO(momernick): Once TensorSpecs are implemented for CTs, use that here.
     inputs = nest.flatten(inputs, expand_composites=True)
 
     session = get_session(inputs)
@@ -3488,10 +3461,8 @@ def __init__(self, inputs, outputs, updates=None, name=None):
               x.op.inputs[0])
 
   def __call__(self, inputs):
-    # Convert scipy sparse data into sparse tensors.
-    inputs = try_convert_scipy_to_sparse(inputs)
-
     input_values = nest.flatten(inputs, expand_composites=True)
+
     if self._freezable_vars_values:
       input_values = input_values + self._freezable_vars_values
     converted_inputs = []

tensorflow/python/keras/engine/input_layer.py

@@ -34,21 +34,27 @@ class InputLayer(base_layer.Layer):
   """Layer to be used as an entry point into a Network (a graph of layers).
 
   It can either wrap an existing tensor (pass an `input_tensor` argument)
-  or create its a placeholder tensor (pass arguments `input_shape`, and
+  or create a placeholder tensor (pass arguments `input_shape`, and
   optionally, `dtype`).
 
   It is generally recommend to use the functional layer API via `Input`,
   (which creates an `InputLayer`) without directly using `InputLayer`.
 
+  This class can create placeholders for tf.Tensors, tf.SparseTensors, and
+  tf.RaggedTensors by choosing 'sparse=True' or 'ragged=True'.
+
   Arguments:
       input_shape: Shape tuple (not including the batch axis), or `TensorShape`
         instance (not including the batch axis).
       batch_size: Optional input batch size (integer or None).
       dtype: Datatype of the input.
       input_tensor: Optional tensor to use as layer input
           instead of creating a placeholder.
-      sparse: Boolean, whether the placeholder created
-          is meant to be sparse.
+      sparse: Boolean, whether the placeholder created is meant to be sparse.
+      ragged: Boolean, whether the placeholder created is meant to be ragged.
+          In this case, values of 'None' in the 'shape' argument represent
+          ragged dimensions. For more information about RaggedTensors, see
+          https://www.tensorflow.org/guide/ragged_tensors.
       name: Name of the layer (string).
   """
 
@@ -59,6 +65,7 @@ def __init__(self,
                input_tensor=None,
                sparse=False,
                name=None,
+               ragged=False,
                **kwargs):
     strategy = distribution_strategy_context.get_strategy()
     if strategy and batch_size is not None and \
@@ -110,18 +117,12 @@ def __init__(self,
         batch_input_shape = None
       graph = backend.get_graph()
       with graph.as_default():
-        # In graph mode, create a graph placeholder to call the layer on.
-        if sparse:
-          input_tensor = backend.placeholder(
-              shape=batch_input_shape,
-              dtype=dtype,
-              name=self.name,
-              sparse=True)
-        else:
-          input_tensor = backend.placeholder(
-              shape=batch_input_shape,
-              dtype=dtype,
-              name=self.name)
+        input_tensor = backend.placeholder(
+            shape=batch_input_shape,
+            dtype=dtype,
+            name=self.name,
+            sparse=sparse,
+            ragged=ragged)
 
       self.is_placeholder = True
       self._batch_input_shape = batch_input_shape
@@ -164,6 +165,7 @@ def Input(  # pylint: disable=invalid-name
     dtype=None,
     sparse=False,
     tensor=None,
+    ragged=False,
     **kwargs):
   """`Input()` is used to instantiate a Keras tensor.
 
@@ -184,17 +186,24 @@ def Input(  # pylint: disable=invalid-name
   Arguments:
       shape: A shape tuple (integers), not including the batch size.
           For instance, `shape=(32,)` indicates that the expected input
-          will be batches of 32-dimensional vectors.
+          will be batches of 32-dimensional vectors. Elements of this tuple
+          can be None; 'None' elements represent dimensions where the shape is
+          not known.
       batch_size: optional static batch size (integer).
       name: An optional name string for the layer.
           Should be unique in a model (do not reuse the same name twice).
           It will be autogenerated if it isn't provided.
       dtype: The data type expected by the input, as a string
           (`float32`, `float64`, `int32`...)
-      sparse: A boolean specifying whether the placeholder
-          to be created is sparse.
+      sparse: A boolean specifying whether the placeholder to be created is
+          sparse. Only one of 'ragged' and 'sparse' can be True.
       tensor: Optional existing tensor to wrap into the `Input` layer.
           If set, the layer will not create a placeholder tensor.
+      ragged: A boolean specifying whether the placeholder to be created is
+          ragged. Only one of 'ragged' and 'sparse' can be True. In this case,
+          values of 'None' in the 'shape' argument represent ragged dimensions.
+          For more information about RaggedTensors, see
+          https://www.tensorflow.org/guide/ragged_tensors.
       **kwargs: deprecated arguments support.
 
   Returns:
@@ -222,6 +231,10 @@ def Input(  # pylint: disable=invalid-name
   Raises:
     ValueError: in case of invalid arguments.
   """
+  if sparse and ragged:
+    raise ValueError(
+        'Cannot set both sparse and ragged to True in a Keras input.')
+
   batch_shape = None
   if 'batch_shape' in kwargs:
     batch_shape = kwargs.pop('batch_shape')
@@ -246,6 +259,7 @@ def Input(  # pylint: disable=invalid-name
         name=name,
         dtype=dtype,
         sparse=sparse,
+        ragged=ragged,
         input_tensor=tensor)
   else:
     input_layer = InputLayer(
@@ -254,6 +268,7 @@ def Input(  # pylint: disable=invalid-name
         name=name,
         dtype=dtype,
         sparse=sparse,
+        ragged=ragged,
         input_tensor=tensor)
 
   # Return tensor including `_keras_history`.