initial commit

theano/sandbox/linalg/__init__.py

@@ -2,5 +2,5 @@
 from kron import kron
 from ops import (cholesky, matrix_inverse, solve,
         diag, extract_diag, alloc_diag,
-        det, psd, eig, eigh,
+        det, psd, eig, eigh, geigvalsh,
         trace, spectral_radius_bound)

theano/sandbox/linalg/ops.py

@@ -1097,3 +1097,108 @@ class EighGrad(Op):
 
     def infer_shape(self, node, shapes):
         return [shapes[0]]
+
+
+
+class GEigvalsh(Op):
+    """Generalized eigenvalues of a Hermetian positive definite eigensystem
+    """
+
+    def __init__(self, lower=True):
+        assert lower in [True, False]
+        self.lower = lower
+
+    def props(self):
+        return (self.lower,)
+
+    def __hash__(self):
+        return hash((type(self), self.props()))
+
+    def __eq__(self, other):
+        return (type(self) == type(other) and self.props() == other.props())
+
+    def make_node(self, a, b):
+        assert imported_scipy, (
+            "Scipy not available. Scipy is needed for the GEigvalsh op")
+        a, b = map(as_tensor_variable, (a, b))
+        assert a.ndim == 2
+        assert b.ndim == 2
+
+        out_dtype = theano.scalar.upcast(a.dtype, b.dtype)
+        w = theano.tensor.vector(dtype=out_dtype)
+        return Apply(self, [a, b], [w])
+
+    def perform(self, node, (a, b), (w,)):
+        try:
+            w[0] = scipy.linalg.eigvalsh(a=a, b=b, lower=self.lower)
+        except numpy.linalg.LinAlgError:
+            logger.debug('Failed to find generalized eigs of %s' % (
+                node.inputs[0]))
+            raise
+
+    def grad(self, inputs, g_outputs):
+        a, b = inputs
+        gw, = g_outputs
+        return GEigvalshGrad(self.lower)(a, b, gw)
+
+    def infer_shape(self, node, shapes):
+        n = shapes[0][0]
+        return [(n,)]
+
+
+class GEigvalshGrad(Op):
+    """Gradient of generalized eigenvalues of a Hermetian positive definite
+    eigensystem
+    """
+
+    def __init__(self, lower=True):
+        assert lower in [True, False]
+        self.lower = lower
+        if lower:
+            self.tri0 = numpy.tril
+            self.tri1 = lambda a: numpy.triu(a, 1)
+        else:
+            self.tri0 = numpy.triu
+            self.tri1 = lambda a: numpy.tril(a, -1)
+
+    def props(self):
+        return (self.lower,)
+
+    def __hash__(self):
+        return hash((type(self), self.props()))
+
+    def __eq__(self, other):
+        return (type(self) == type(other) and self.props() == other.props())
+
+    def make_node(self, a, b, gw):
+        assert imported_scipy, (
+            "Scipy not available. Scipy is needed for the GEigvalsh op")
+        a, b, gw = map(as_tensor_variable, (a, b, gw))
+        assert a.ndim == 2
+        assert b.ndim == 2
+        assert gw.ndim == 1
+
+        out_dtype = theano.scalar.upcast(a.dtype, b.dtype, gw.dtype)
+        out1 = theano.tensor.matrix(dtype=out_dtype)
+        out2 = theano.tensor.matrix(dtype=out_dtype)
+        return Apply(self, [a, b, gw], [out1, out2])
+
+    def perform(self, node, (a, b, gw), outputs):
+        N = a.shape[0]
+        w, v = scipy.linalg.eigh(a, b, lower=True)
+
+        gA = v.dot(numpy.diag(gw).dot(v.T))
+        gB = - v.dot(numpy.diag(gw*w).dot(v.T))
+
+        # See EighGrad comments for an explanation of these lines
+        out1 = self.tri0(gA) + self.tri1(gA).T
+        out2 = self.tri0(gB) + self.tri1(gB).T
+        outputs[0][0] = numpy.asarray(out1, dtype=node.outputs[0].dtype)
+        outputs[1][0] = numpy.asarray(out2, dtype=node.outputs[1].dtype)
+
+    def infer_shape(self, node, shapes):
+         return [shapes[0], shapes[1]]
+
+
+def geigvalsh(a, b, lower=True):
+     return GEigvalsh(lower)(a, b)

theano/sandbox/linalg/tests/test_linalg.py

@@ -32,7 +32,7 @@ from theano.sandbox.linalg.ops import (cholesky,
                                        Eig,
                                        inv_as_solve,
                                        )
-from theano.sandbox.linalg import eig, eigh
+from theano.sandbox.linalg import eig, eigh, geigvalsh
 from nose.plugins.skip import SkipTest
 from nose.plugins.attrib import attr
 
@@ -573,3 +573,31 @@ def test_matrix_inverse_solve():
     node = matrix_inverse(A).dot(b).owner
     [out] = inv_as_solve.transform(node)
     assert isinstance(out.owner.op, Solve)
+
+
+def test_geigvalsh():
+    if not imported_scipy:
+        raise SkipTest("Scipy needed for the Solve op.")
+    import scipy.linalg
+
+    A = theano.tensor.dmatrix('a')
+    B = theano.tensor.dmatrix('b')
+    f = function([A,B], geigvalsh(A, B))
+
+    rng = numpy.random.RandomState(utt.fetch_seed())
+    a = rng.randn(5, 5)
+    a = a + a.T
+    b = 10 * numpy.eye(5, 5) + rng.randn(5, 5)
+
+    w = f(a,b)
+    refw = scipy.linalg.eigvalsh(a, b)
+    numpy.testing.assert_array_almost_equal(w, refw)
+
+
+def test_geigvalsh_grad():
+    rng = numpy.random.RandomState(utt.fetch_seed())
+    a = rng.randn(5, 5)
+    a = a + a.T
+    b = 10 * numpy.eye(5, 5) + rng.randn(5, 5)
+    tensor.verify_grad(lambda a, b: geigvalsh(a, b).dot([1,2,3,4,5]), [a, b],
+        rng=numpy.random)