matrix_adjoint added to contrib/linalg/linear_operator_util

Change: 152404828

4 files changed, 82 insertions, 1 deletions

diff --git a/tensorflow/contrib/linalg/python/kernel_tests/linear_operator_util_test.py b/tensorflow/contrib/linalg/python/kernel_tests/linear_operator_util_test.py
index a06af336e7..f047f4b978 100644
--- a/tensorflow/contrib/linalg/python/kernel_tests/linear_operator_util_test.py
+++ b/tensorflow/contrib/linalg/python/kernel_tests/linear_operator_util_test.py
@@ -229,6 +229,29 @@ class MatmulWithBroadcastTest(test.TestCase):
       self.assertAllEqual(expected, result)
 
 
+class MatrixAdjointTest(test.TestCase):
+
+  def testNonBatchMatrix(self):
+    a = [[1, 2, 3j], [4, 5, -6j]]  # Shape (2, 3)
+    expected = [[1, 4], [2, 5], [-3j, 6j]]  # Shape (3, 2)
+    with self.test_session():
+      a_adj = linear_operator_util.matrix_adjoint(a)
+      self.assertEqual((3, 2), a_adj.get_shape())
+      self.assertAllClose(expected, a_adj.eval())
+
+  def testBatchMatrix(self):
+    matrix_0 = [[1j, 2, 3], [4, 5, 6]]
+    matrix_0_a = [[-1j, 4], [2, 5], [3, 6]]
+    matrix_1 = [[11, 22, 33], [44, 55, 66j]]
+    matrix_1_a = [[11, 44], [22, 55], [33, -66j]]
+    batch_matrix = [matrix_0, matrix_1]  # Shape (2, 2, 3)
+    expected_adj = [matrix_0_a, matrix_1_a]  # Shape (2, 3, 2)
+    with self.test_session():
+      matrix_adj = linear_operator_util.matrix_adjoint(batch_matrix)
+      self.assertEqual((2, 3, 2), matrix_adj.get_shape())
+      self.assertAllEqual(expected_adj, matrix_adj.eval())
+
+
 class DomainDimensionStubOperator(object):
 
   def __init__(self, domain_dimension):
diff --git a/tensorflow/contrib/linalg/python/ops/linear_operator_util.py b/tensorflow/contrib/linalg/python/ops/linear_operator_util.py
index a52a235677..9f8cb23169 100644
--- a/tensorflow/contrib/linalg/python/ops/linear_operator_util.py
+++ b/tensorflow/contrib/linalg/python/ops/linear_operator_util.py
@@ -289,6 +289,53 @@ def matmul_with_broadcast(a,
         b_is_sparse=b_is_sparse)
 
 
+def matrix_adjoint(a, name="matrix_adjoint"):
+  """Transposes last two dimensions of tensor `a`, and takes complex conjugate.
+
+  If `a` is real valued, the result is equivalent to `matrix_transpose`.
+
+  For example:
+
+  ```python
+  # Matrix with no batch dimension.
+  # 'x' is [[1 2 3j]
+  #         [4 5 -6j]]
+  tf.matrix_adjoint(x) ==> [[1 4]
+                            [2 5]
+                            [-3j 6j]]
+
+  # Matrix with two batch dimensions.
+  # x.shape is [1, 2, 3, 4]
+  # tf.matrix_adjoint(x) is shape [1, 2, 4, 3]
+  ```
+
+  Note that `tf.matmul` provides kwargs allowing for adjoint of arguments.  This
+  is done with minimal cost, and is preferable to using this function. E.g.
+
+  ```
+  # Good!  Adjoint is taken at minimal additional cost.
+  tf.matmul(matrix, b, adjoint_b=True)
+
+  # Inefficient!
+  tf.matmul(matrix, tf.matrix_adjoint(b))
+  ```
+
+  Args:
+    a: A `Tensor` with `rank >= 2`.
+    name: A name for the operation (optional).
+
+  Returns:
+    A batch matrix `Tensor` with same `dtype` as `a`.
+
+  Raises:
+    ValueError:  If `a` is determined statically to have `rank < 2`.
+  """
+  with ops.name_scope(name, values=[a]):
+    a = ops.convert_to_tensor(a, name="a")
+    a_transpose = array_ops.matrix_transpose(a)
+    return math_ops.conj(a_transpose)
+
+
 def shape_tensor(shape, name=None):
   """Convert Tensor using default type, unless empty list or tuple."""
   # Works just like random_ops._ShapeTensor.
diff --git a/tensorflow/python/kernel_tests/linalg_ops_test.py b/tensorflow/python/kernel_tests/linalg_ops_test.py
index ff299e6511..153d4ab662 100644
--- a/tensorflow/python/kernel_tests/linalg_ops_test.py
+++ b/tensorflow/python/kernel_tests/linalg_ops_test.py
@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-"""Tests for tensorflow.python.ops.special_math_ops."""
+"""Tests for tensorflow.python.ops.linalg_ops."""
 
 from __future__ import absolute_import
 from __future__ import division
diff --git a/tensorflow/python/ops/array_ops.py b/tensorflow/python/ops/array_ops.py
index 60057b9ab1..5ebfe5b89f 100644
--- a/tensorflow/python/ops/array_ops.py
+++ b/tensorflow/python/ops/array_ops.py
@@ -1292,6 +1292,17 @@ def matrix_transpose(a, name="matrix_transpose"):
   # tf.matrix_transpose(x) is shape [1, 2, 4, 3]
   ```
 
+  Note that `tf.matmul` provides kwargs allowing for transpose of arguments.
+  This is done with minimal cost, and is preferable to using this function. E.g.
+
+  ```
+  # Good!  Transpose is taken at minimal additional cost.
+  tf.matmul(matrix, b, transpose_b=True)
+
+  # Inefficient!
+  tf.matmul(matrix, tf.matrix_transpose(b))
+  ```
+
   Args:
     a: A `Tensor` with `rank >= 2`.
     name: A name for the operation (optional).