test passing

theano/gpuarray/subtensor.py

@@ -1102,9 +1102,7 @@ class GpuExtractDiag(Subtensor):
         broadcastable = x.broadcastable[:axis_small] + \
             x.broadcastable[axis_small + 1:axis_large] + \
             x.broadcastable[axis_large + 1:] + (False,)
-        return gof.Apply(self, [x], [x.type.__class__(
-            dtype=x.dtype,
-            broadcastable=broadcastable)()])
+        return gof.Apply(self, [x], [x.type.clone(broadcastable=broadcastable)()])
 
     def perform(self, node, inputs, outputs):
         (x,) = inputs
@@ -1197,9 +1195,8 @@ class GpuAllocDiag(Op):
 
         if x.ndim != 1:
             raise ValueError('AllocDiag argument must be a vector!', x)
-        
-        return gof.Apply(self, [x], [x.type.__class__(dtype=x.dtype,
-            broadcastable=x.broadcastable)()])
+
+        return gof.Apply(self, [x], [x.type.clone(broadcastable=(False, False))()])
 
     def perform(self, node, inputs, outputs):
         (x,) = inputs
@@ -1218,7 +1215,7 @@ class GpuAllocDiag(Op):
 
     def grad(self, inputs, gout):
         (gz,) = gout
-        return [GpuExtractDiag(offset=self.offset, axis1=0, axis2=1, view=False)(gz)]
+        return [GpuExtractDiag(offset=self.offset, axis1=0, axis2=1)(gz)]
 
     def infer_shape(self, node, shapes):
         dim = shapes[0][0] + abs(self.offset)

theano/gpuarray/tests/test_subtensor.py

@@ -177,7 +177,7 @@ def test_adv_subtensor():
     assert np.allclose(rval, rep)
 
 
-class test_gpudiagonal(unittest.TestCase):
+class test_gpuextractdiag(unittest.TestCase):
     def test_matrix(self):
         x = tensor.matrix()
         np_x = np.arange(77).reshape(7, 11).astype(theano.config.floatX)
@@ -213,8 +213,48 @@ class test_gpuallocdiag(unittest.TestCase):
 
     def test_grad(self):
         x = tensor.vector()
-        np_x = np.arange(7).astype(theano.config.floatX)
+        np_x = np.random.randn(7).astype(theano.config.floatX)
+        
+        # offset = 0 case:
         mtx_x = GpuAllocDiag()(x)
         sum_mtx_x = tensor.sum(mtx_x)
-        grad = tensor.grad(sum_mtx_x, x)
+        grad_x = tensor.grad(sum_mtx_x, x)
+        grad_mtx_x = tensor.grad(sum_mtx_x, mtx_x)
+
+        fn_grad_x = theano.function([x], grad_x)
+        fn_grad_mtx_x = theano.function([x], grad_mtx_x)
+
+        computed_grad_x = fn_grad_x(np_x)
+        computed_grad_mtx_x = fn_grad_mtx_x(np_x)
+        true_grad_x = np.diagonal(computed_grad_mtx_x, 0)
+        assert np.allclose(computed_grad_x, true_grad_x)
+
+        # offset > 0 case:
+        mtx_x = GpuAllocDiag(2)(x)
+        sum_mtx_x = tensor.sum(mtx_x)
+        grad_x = tensor.grad(sum_mtx_x, x)
+        grad_mtx_x = tensor.grad(sum_mtx_x, mtx_x)
+
+        fn_grad_x = theano.function([x], grad_x)
+        fn_grad_mtx_x = theano.function([x], grad_mtx_x)
+
+        computed_grad_x = fn_grad_x(np_x)
+        computed_grad_mtx_x = fn_grad_mtx_x(np_x)
+        true_grad_x = np.diagonal(computed_grad_mtx_x, 2)
+        assert np.allclose(computed_grad_x, true_grad_x)
+
+        # offset < 0 case:
+        mtx_x = GpuAllocDiag(-3)(x)
+        sum_mtx_x = tensor.sum(mtx_x)
+        grad_x = tensor.grad(sum_mtx_x, x)
+        grad_mtx_x = tensor.grad(sum_mtx_x, mtx_x)
+
+        fn_grad_x = theano.function([x], grad_x)
+        fn_grad_mtx_x = theano.function([x], grad_mtx_x)
+
+        computed_grad_x = fn_grad_x(np_x)
+        computed_grad_mtx_x = fn_grad_mtx_x(np_x)
+        true_grad_x = np.diagonal(computed_grad_mtx_x, -3)
+        assert np.allclose(computed_grad_x, true_grad_x)
+
         # assert

theano/tensor/basic.py

@@ -6167,6 +6167,24 @@ class ExtractDiag(Op):
 
 
 def diagonal(a, offset=0, axis1=0, axis2=1):
+    """
+    A helper function for `theano.tensor.ExtractDiag`. It accepts tensor with
+    `ndim >= 2` as input. The name `diagonal` is just meant to keep it
+    consistent with numpy.
+
+    Parameters
+    ----------
+    a : symbolic tensor
+    offset : int
+        offset
+    axis1 : int
+    axis2 : int
+
+    Returns
+    -------
+    tensor : symbolic tensor
+
+    """
     return ExtractDiag(offset, axis1, axis2)(a)
 
 
@@ -6223,12 +6241,13 @@ class AllocDiag(Op):
 
 def diag(v, k=0):
     """
-    A helper function for two ops: theano.tensor.ExtractDiag and
-    theano.tensor.AllocDiag. It both accepts tensor vector and tensor matrix.
-     While the passed tensor variable `v` has `v.ndim>=2`, it builds a
-     ExtractDiag instance, and returns a vector with its entries equal to
-     `v`'s main diagonal; otherwise if `v.ndim` is `1`, it builds a AllocDiag
-     instance, and returns a matrix with `v` at its k-th diaogonal.
+    A helper function for two ops: `theano.tensor.ExtractDiag` and
+    `theano.tensor.AllocDiag`. The name `diag` is meant to keep it consistent
+    with numpy. It both accepts tensor vector and tensor matrix.
+    While the passed tensor variable `v` has `v.ndim>=2`, it builds a
+    `ExtractDiag` instance, and returns a vector with its entries equal to
+    `v`'s main diagonal; otherwise if `v.ndim` is `1`, it builds an `AllocDiag`
+    instance, and returns a matrix with `v` at its k-th diaogonal.
 
     Parameters
     ----------