Deprecate C++ kernel for matrix exponential, which is now implemented as a python function.

PiperOrigin-RevId: 210180168

tensorflow/core/api_def/base_api/api_def_MatrixExponential.pbtxt

@@ -1,32 +1,5 @@
 op {
   graph_op_name: "MatrixExponential"
-  in_arg {
-    name: "input"
-    description: <<END
-Shape is `[..., M, M]`.
-END
-  }
-  out_arg {
-    name: "output"
-    description: <<END
-Shape is `[..., M, M]`.
-
-@compatibility(scipy)
-Equivalent to scipy.linalg.expm
-@end_compatibility
-END
-  }
-  summary: "Computes the matrix exponential of one or more square matrices:"
-  description: <<END
-\\(exp(A) = \sum_{n=0}^\infty A^n/n!\\)
-
-The exponential is computed using a combination of the scaling and squaring
-method and the Pade approximation. Details can be founds in:
-Nicholas J. Higham, "The scaling and squaring method for the matrix exponential
-revisited," SIAM J. Matrix Anal. Applic., 26:1179-1193, 2005.
-
-The input is a tensor of shape `[..., M, M]` whose inner-most 2 dimensions
-form square matrices. The output is a tensor of the same shape as the input
-containing the exponential for all input submatrices `[..., :, :]`.
-END
+  visibility: SKIP
+  summary: "Deprecated, use python implementation tf.linalg.matrix_exponential."
 }

tensorflow/core/kernels/matrix_exponential_op.cc

@@ -49,6 +49,7 @@ class MatrixExponentialOp : public LinearAlgebraOp<Scalar> {
   TF_DISALLOW_COPY_AND_ASSIGN(MatrixExponentialOp);
 };
 
+// Deprecated kernels (2018/08/21).
 REGISTER_LINALG_OP("MatrixExponential", (MatrixExponentialOp<float>), float);
 REGISTER_LINALG_OP("MatrixExponential", (MatrixExponentialOp<double>), double);
 REGISTER_LINALG_OP("MatrixExponential", (MatrixExponentialOp<complex64>),

tensorflow/core/ops/linalg_ops.cc

@@ -235,6 +235,8 @@ REGISTER_OP("MatrixInverse")
     .SetShapeFn(BatchUnchangedSquareShapeFn);
 
 REGISTER_OP("MatrixExponential")
+    .Deprecated(
+        27, "Use Python implementation tf.linalg.matrix_exponential instead.")
     .Input("input: T")
     .Output("output: T")
     .Attr("T: {double, float, complex64, complex128}")

tensorflow/core/public/version.h

@@ -96,10 +96,12 @@ limitations under the License.
 //     GraphDef. (7dec2017)
 // 27. Deprecate TensorArray ops v2 in favor of v3 and deprecated io_ops
 //     deprecated in favor of V2 ops. (2018/01/23)
+// 28. Deprecate MatrixExponential op in favor of Python implementation.
+//     (2018/08/21).
 
 #define TF_GRAPH_DEF_VERSION_MIN_PRODUCER 0
 #define TF_GRAPH_DEF_VERSION_MIN_CONSUMER 0
-#define TF_GRAPH_DEF_VERSION 26
+#define TF_GRAPH_DEF_VERSION 27
 
 // Checkpoint compatibility versions (the versions field in SavedSliceMeta).
 //

tensorflow/python/kernel_tests/BUILD

@@ -601,7 +601,7 @@ tf_py_test(
 
 tf_py_test(
     name = "matrix_logarithm_op_test",
-    size = "small",
+    size = "medium",
     srcs = ["matrix_logarithm_op_test.py"],
     additional_deps = [
         "//third_party/py/numpy",

tensorflow/python/kernel_tests/matrix_logarithm_op_test.py

@@ -30,6 +30,7 @@ from tensorflow.python.ops import gen_linalg_ops
 from tensorflow.python.ops import math_ops
 from tensorflow.python.ops import random_ops
 from tensorflow.python.ops import variables
+from tensorflow.python.ops.linalg import linalg_impl
 from tensorflow.python.platform import test
 
 
@@ -39,7 +40,7 @@ class LogarithmOpTest(test.TestCase):
     inp = x.astype(np_type)
     with self.test_session(use_gpu=True):
       # Verify that expm(logm(A)) == A.
-      tf_ans = gen_linalg_ops.matrix_exponential(
+      tf_ans = linalg_impl.matrix_exponential(
           gen_linalg_ops.matrix_logarithm(inp))
       out = tf_ans.eval()
       self.assertAllClose(inp, out, rtol=1e-4, atol=1e-3)
@@ -98,16 +99,25 @@ class LogarithmOpTest(test.TestCase):
     self._verifyLogarithmComplex(np.empty([0, 2, 2], dtype=np.complex64))
     self._verifyLogarithmComplex(np.empty([2, 0, 0], dtype=np.complex64))
 
-  def testRandomSmallAndLarge(self):
+  def testRandomSmallAndLargeComplex64(self):
     np.random.seed(42)
-    for dtype in np.complex64, np.complex128:
-      for batch_dims in [(), (1,), (3,), (2, 2)]:
-        for size in 8, 31, 32:
-          shape = batch_dims + (size, size)
-          matrix = np.random.uniform(
-              low=-1.0, high=1.0,
-              size=np.prod(shape)).reshape(shape).astype(dtype)
-          self._verifyLogarithmComplex(matrix)
+    for batch_dims in [(), (1,), (3,), (2, 2)]:
+      for size in 8, 31, 32:
+        shape = batch_dims + (size, size)
+        matrix = np.random.uniform(
+            low=-1.0, high=1.0,
+            size=np.prod(shape)).reshape(shape).astype(np.complex64)
+        self._verifyLogarithmComplex(matrix)
+
+  def testRandomSmallAndLargeComplex128(self):
+    np.random.seed(42)
+    for batch_dims in [(), (1,), (3,), (2, 2)]:
+      for size in 8, 31, 32:
+        shape = batch_dims + (size, size)
+        matrix = np.random.uniform(
+            low=-1.0, high=1.0,
+            size=np.prod(shape)).reshape(shape).astype(np.complex128)
+        self._verifyLogarithmComplex(matrix)
 
   def testConcurrentExecutesWithoutError(self):
     with self.test_session(use_gpu=True) as sess: