Python linalg_ops.self_adjoint_eig函数代码示例

 def testWrongDimensions(self):
   # The input to self_adjoint_eig should be a tensor of
   # at least rank 2.
   scalar = constant_op.constant(1.)
   with self.assertRaises(ValueError):
     linalg_ops.self_adjoint_eig(scalar)
   vector = constant_op.constant([1., 2.])
   with self.assertRaises(ValueError):
     linalg_ops.self_adjoint_eig(vector)

  def Test(self):
    np.random.seed(1)
    n = shape_[-1]
    batch_shape = shape_[:-2]
    np_dtype = dtype_.as_numpy_dtype
    a = np.random.uniform(
        low=-1.0, high=1.0, size=n * n).reshape([n, n]).astype(np_dtype)
    if dtype_.is_complex:
      a += 1j * np.random.uniform(
          low=-1.0, high=1.0, size=n * n).reshape([n, n]).astype(np_dtype)
    a += np.conj(a.T)
    a = np.tile(a, batch_shape + (1, 1))
    if dtype_ in (dtypes_lib.float32, dtypes_lib.complex64):
      atol = 1e-4
    else:
      atol = 1e-12
    np_e, np_v = np.linalg.eigh(a)
    with self.test_session():
      if compute_v_:
        tf_e, tf_v = linalg_ops.self_adjoint_eig(constant_op.constant(a))

        # Check that V*diag(E)*V^T is close to A.
        a_ev = math_ops.matmul(
            math_ops.matmul(tf_v, array_ops.matrix_diag(tf_e)),
            tf_v,
            adjoint_b=True)
        self.assertAllClose(a_ev.eval(), a, atol=atol)

        # Compare to numpy.linalg.eigh.
        CompareEigenDecompositions(self, np_e, np_v,
                                   tf_e.eval(), tf_v.eval(), atol)
      else:
        tf_e = linalg_ops.self_adjoint_eigvals(constant_op.constant(a))
        self.assertAllClose(
            np.sort(np_e, -1), np.sort(tf_e.eval(), -1), atol=atol)

 def Test(self):
   np.random.seed(1)
   n = shape_[-1]
   batch_shape = shape_[:-2]
   a = np.random.uniform(
       low=-1.0, high=1.0, size=n * n).reshape([n, n]).astype(dtype_)
   a += np.conj(a.T)
   a = np.tile(a, batch_shape + (1, 1))
   # Optimal stepsize for central difference is O(epsilon^{1/3}).
   epsilon = np.finfo(dtype_).eps
   delta = 0.1 * epsilon**(1.0 / 3.0)
   # tolerance obtained by looking at actual differences using
   # np.linalg.norm(theoretical-numerical, np.inf) on -mavx build
   if dtype_ == np.float32:
     tol = 1e-2
   else:
     tol = 1e-7
   with self.test_session():
     tf_a = constant_op.constant(a)
     tf_e, tf_v = linalg_ops.self_adjoint_eig(tf_a)
     for b in tf_e, tf_v:
       x_init = np.random.uniform(
           low=-1.0, high=1.0, size=n * n).reshape([n, n]).astype(dtype_)
       x_init += np.conj(x_init.T)
       x_init = np.tile(x_init, batch_shape + (1, 1))
       theoretical, numerical = gradient_checker.compute_gradient(
           tf_a,
           tf_a.get_shape().as_list(),
           b,
           b.get_shape().as_list(),
           x_init_value=x_init,
           delta=delta)
       self.assertAllClose(theoretical, numerical, atol=tol, rtol=tol)

 def Test(self):
   np.random.seed(1)
   n = shape_[-1]
   batch_shape = shape_[:-2]
   np_dtype = dtype_.as_numpy_dtype
   a = np.random.uniform(
       low=-1.0, high=1.0, size=n * n).reshape([n, n]).astype(np_dtype)
   if dtype_.is_complex:
     a += 1j * np.random.uniform(
         low=-1.0, high=1.0, size=n * n).reshape([n, n]).astype(np_dtype)
   a += np.conj(a.T)
   a = np.tile(a, batch_shape + (1, 1))
   # Optimal stepsize for central difference is O(epsilon^{1/3}).
   epsilon = np.finfo(np_dtype).eps
   delta = 0.1 * epsilon**(1.0 / 3.0)
   # tolerance obtained by looking at actual differences using
   # np.linalg.norm(theoretical-numerical, np.inf) on -mavx build
   if dtype_ in (dtypes_lib.float32, dtypes_lib.complex64):
     tol = 1e-2
   else:
     tol = 1e-7
   with self.session(use_gpu=True):
     tf_a = constant_op.constant(a)
     if compute_v_:
       tf_e, tf_v = linalg_ops.self_adjoint_eig(tf_a)
       # (complex) Eigenvectors are only unique up to an arbitrary phase
       # We normalize the vectors such that the first component has phase 0.
       top_rows = tf_v[..., 0:1, :]
       if tf_a.dtype.is_complex:
         angle = -math_ops.angle(top_rows)
         phase = math_ops.complex(math_ops.cos(angle), math_ops.sin(angle))
       else:
         phase = math_ops.sign(top_rows)
       tf_v *= phase
       outputs = [tf_e, tf_v]
     else:
       tf_e = linalg_ops.self_adjoint_eigvals(tf_a)
       outputs = [tf_e]
     for b in outputs:
       x_init = np.random.uniform(
           low=-1.0, high=1.0, size=n * n).reshape([n, n]).astype(np_dtype)
       if dtype_.is_complex:
         x_init += 1j * np.random.uniform(
             low=-1.0, high=1.0, size=n * n).reshape([n, n]).astype(np_dtype)
       x_init += np.conj(x_init.T)
       x_init = np.tile(x_init, batch_shape + (1, 1))
       theoretical, numerical = gradient_checker.compute_gradient(
           tf_a,
           tf_a.get_shape().as_list(),
           b,
           b.get_shape().as_list(),
           x_init_value=x_init,
           delta=delta)
       self.assertAllClose(theoretical, numerical, atol=tol, rtol=tol)

 def testConcurrentExecutesWithoutError(self):
   all_ops = []
   with self.session(use_gpu=True) as sess:
     for compute_v_ in True, False:
       matrix1 = random_ops.random_normal([5, 5], seed=42)
       matrix2 = random_ops.random_normal([5, 5], seed=42)
       if compute_v_:
         e1, v1 = linalg_ops.self_adjoint_eig(matrix1)
         e2, v2 = linalg_ops.self_adjoint_eig(matrix2)
         all_ops += [e1, v1, e2, v2]
       else:
         e1 = linalg_ops.self_adjoint_eigvals(matrix1)
         e2 = linalg_ops.self_adjoint_eigvals(matrix2)
         all_ops += [e1, e2]
     val = sess.run(all_ops)
     self.assertAllEqual(val[0], val[2])
     # The algorithm is slightly different for compute_v being True and False,
     # so require approximate equality only here.
     self.assertAllClose(val[2], val[4])
     self.assertAllEqual(val[4], val[5])
     self.assertAllEqual(val[1], val[3])

 def Compute(x):
   e, v = linalg_ops.self_adjoint_eig(x)
   # (complex) Eigenvectors are only unique up to an arbitrary phase
   # We normalize the vectors such that the first component has phase 0.
   top_rows = v[..., 0:1, :]
   if dtype_.is_complex:
     angle = -math_ops.angle(top_rows)
     phase = math_ops.complex(math_ops.cos(angle), math_ops.sin(angle))
   else:
     phase = math_ops.sign(top_rows)
   v *= phase
   return e, v

 def testMatrixThatFailsWhenFlushingDenormsToZero(self):
   # Test a 32x32 matrix which is known to fail if denorm floats are flushed to
   # zero.
   matrix = np.genfromtxt(
       test.test_src_dir_path(
           "python/kernel_tests/testdata/"
           "self_adjoint_eig_fail_if_denorms_flushed.txt")).astype(np.float32)
   self.assertEqual(matrix.shape, (32, 32))
   matrix_tensor = constant_op.constant(matrix)
   with self.session(use_gpu=True) as sess:
     (e, v) = sess.run(linalg_ops.self_adjoint_eig(matrix_tensor))
     self.assertEqual(e.size, 32)
     self.assertAllClose(
         np.matmul(v, v.transpose()), np.eye(32, dtype=np.float32), atol=2e-3)
     self.assertAllClose(matrix,
                         np.matmul(np.matmul(v, np.diag(e)), v.transpose()))

  def get_eigendecomp(self):
    """Creates or retrieves eigendecomposition of self._cov."""
    # Unlike get_inverse and get_matpower this doesn't retrieve a stored
    # variable, but instead always computes a fresh version from the current
    # value of get_cov().
    if not self._eigendecomp:
      eigenvalues, eigenvectors = linalg_ops.self_adjoint_eig(self._cov)

      # The matrix self._cov is positive semidefinite by construction, but the
      # numerical eigenvalues could be negative due to numerical errors, so here
      # we clip them to be at least FLAGS.eigenvalue_clipping_threshold
      clipped_eigenvalues = math_ops.maximum(eigenvalues,
                                             EIGENVALUE_CLIPPING_THRESHOLD)
      self._eigendecomp = (clipped_eigenvalues, eigenvectors)

    return self._eigendecomp

  def _test(self, dtype, shape):
    np.random.seed(1)
    x_np = np.random.uniform(
        low=-1.0, high=1.0, size=np.prod(shape)).reshape(shape).astype(dtype)
    x_np = x_np + np.swapaxes(x_np, -1, -2)
    n = shape[-1]

    e_np, _ = np.linalg.eigh(x_np)
    with self.cached_session() as sess:
      x_tf = array_ops.placeholder(dtype)
      with self.test_scope():
        e, v = linalg_ops.self_adjoint_eig(x_tf)
      e_val, v_val = sess.run([e, v], feed_dict={x_tf: x_np})

      v_diff = np.matmul(v_val, np.swapaxes(v_val, -1, -2)) - np.eye(n)
      self.assertAlmostEqual(np.mean(v_diff**2), 0.0, delta=1e-6)
      self.assertAlmostEqual(np.mean((e_val - e_np)**2), 0.0, delta=1e-6)

  def register_eigendecomp(self):
    """Registers an eigendecomposition.

    Unlike register_damp_inverse and register_matpower this doesn't create
    any variables or inverse ops.  Instead it merely makes tensors containing
    the eigendecomposition available to anyone that wants them.  They will be
    recomputed (once) for each session.run() call (when they needed by some op).
    """
    if not self._eigendecomp:
      eigenvalues, eigenvectors = linalg_ops.self_adjoint_eig(self._cov)

      # The matrix self._cov is positive semidefinite by construction, but the
      # numerical eigenvalues could be negative due to numerical errors, so here
      # we clip them to be at least FLAGS.eigenvalue_clipping_threshold
      clipped_eigenvalues = math_ops.maximum(eigenvalues,
                                             EIGENVALUE_CLIPPING_THRESHOLD)
      self._eigendecomp = (clipped_eigenvalues, eigenvectors)

def _SelfAdjointEigV2Grad(op, grad_e, grad_v):
  """Gradient for SelfAdjointEigV2."""
  e = op.outputs[0]
  compute_v = op.get_attr("compute_v")
  # a = op.inputs[0], which satisfies
  # a[...,:,:] * v[...,:,i] = e[...,i] * v[...,i]
  with ops.control_dependencies([grad_e, grad_v]):
    if compute_v:
      v = op.outputs[1]
      # Construct the matrix f(i,j) = (i != j ? 1 / (e_i - e_j) : 0).
      # Notice that because of the term involving f, the gradient becomes
      # infinite (or NaN in practice) when eigenvalues are not unique.
      # Mathematically this should not be surprising, since for (k-fold)
      # degenerate eigenvalues, the corresponding eigenvectors are only defined
      # up to arbitrary rotation in a (k-dimensional) subspace.
      f = array_ops.matrix_set_diag(
          math_ops.reciprocal(
              array_ops.expand_dims(e, -2) - array_ops.expand_dims(e, -1)),
          array_ops.zeros_like(e))
      grad_a = math_ops.matmul(
          v,
          math_ops.matmul(
              array_ops.matrix_diag(grad_e) +
              f * math_ops.matmul(v, grad_v, adjoint_a=True),
              v,
              adjoint_b=True))
    else:
      _, v = linalg_ops.self_adjoint_eig(op.inputs[0])
      grad_a = math_ops.matmul(v,
                               math_ops.matmul(
                                   array_ops.matrix_diag(grad_e),
                                   v,
                                   adjoint_b=True))
    # The forward op only depends on the lower triangular part of a, so here we
    # symmetrize and take the lower triangle
    grad_a = array_ops.matrix_band_part(
        grad_a + math_ops.conj(array_ops.matrix_transpose(grad_a)), -1, 0)
    grad_a = array_ops.matrix_set_diag(grad_a,
                                       0.5 * array_ops.matrix_diag_part(grad_a))
    return grad_a

 def register_eigendecomp(self):
   """Registers that an eigendecomposition is needed by a FisherBlock."""
   if not self._eigendecomp:
     self._eigendecomp = linalg_ops.self_adjoint_eig(self._cov)

def posdef_inv_eig(tensor, identity, damping):
  """Computes inverse(tensor + damping * identity) with eigendecomposition."""
  eigenvalues, eigenvectors = linalg_ops.self_adjoint_eig(
      tensor + damping * identity)
  return math_ops.matmul(
      eigenvectors / eigenvalues, eigenvectors, transpose_b=True)

def posdef_eig_self_adjoint(mat):
  """Computes eigendecomposition using self_adjoint_eig."""
  evals, evecs = linalg_ops.self_adjoint_eig(mat)
  evals = math_ops.abs(evals)  # Should be equivalent to svd approach.

  return evals, evecs