Merge pull request #1264 from nouiz/sparse

theano/sparse/basic.py

@@ -3152,12 +3152,24 @@ class Dot(gof.op.Op):
 
         if not x_is_sparse_var:
             x = tensor.as_tensor_variable(x)
+            if x.ndim not in (1, 2):
+                raise TypeError(
+                    'theano.sparse.Dot: input 0 (0-indexed) must have ndim of '
+                    '1 or 2, %d given.' % x.ndim)
 
         if not y_is_sparse_var:
             y = tensor.as_tensor_variable(y)
+            if y.ndim not in (1, 2):
+                raise TypeError(
+                    'theano.sparse.Dot: input 1 (1-indexed) must have ndim of '
+                    '1 or 2, %d given.' % y.ndim)
 
+        if y.ndim == 1 or x.ndim == 1:
+            bz = (False,)
+        else:
+            bz = (False, False)
         return gof.Apply(self, [x, y], [tensor.tensor(dtype=dtype_out,
-                         broadcastable=(False, False))])
+                                                      broadcastable=bz)])
 
     def perform(self, node, inputs, out):
         x, y = inputs
@@ -3294,7 +3306,10 @@ usmm = Usmm()
 
 
 class ConstructSparseFromList(gof.Op):
-    """Constructs a sparse matrix out of a list of 2-D matrix rows"""
+    """Constructs a sparse matrix out of a list of 2-D matrix rows
+
+    :note: The grad implemented is regular, i.e. not structured.
+    """
     def __hash__(self):
         return hash((type(self)))
 
@@ -3304,33 +3319,59 @@ class ConstructSparseFromList(gof.Op):
     def __str__(self):
         return self.__class__.__name__
 
-    def make_node(self, x, y, ilist):
+    def make_node(self, x, values, ilist):
+        """
+        :param x: a dense matrix that specify the output shape.
+        :param values: a dense matrix with the values to use for output.
+        :param ilist: a dense vector with the same lenght as the number of rows
+                      then values. It specify where in the output to put
+                      the corresponding rows.
+
+        This create a sparse matrix with the same shape as `x`. Its
+        values are the rows of `values` moved. Pseudo-code::
+
+            output = csc_matrix.zeros_like(x, dtype=values.dtype)
+            for in_idx, out_idx in enumerate(ilist):
+                output[out_idx] = values[in_idx]
+
+        """
         x_ = theano.tensor.as_tensor_variable(x)
-        y_ = theano.tensor.as_tensor_variable(y)
+        values_ = theano.tensor.as_tensor_variable(values)
         ilist_ = theano.tensor.as_tensor_variable(ilist)
 
         if ilist_.type.dtype[:3] not in ('int', 'uin'):
             raise TypeError('index must be integers')
         if ilist_.type.ndim != 1:
             raise TypeError('index must be vector')
-        if x_.type.ndim == 0:
-            raise TypeError('cannot index into a scalar')
-        if y_.type.ndim > x_.type.ndim:
-            raise TypeError('cannot construct sparse matrix as dimensions differ')    
-        return gof.Apply(self, [x_, y_, ilist_], [theano.sparse.csc_matrix(dtype=x.dtype)])
+        if x_.type.ndim != 2:
+            raise TypeError(
+                'cannot create a sparse matrix with %d dimensions' %
+                x_.type.ndim)
+        if values_.type.ndim != 2:
+            raise TypeError(
+                'cannot create a sparse matrix from values with %d ndim' %
+                values_.type.ndim)
+
+        # We only need the shape of `x` in the perform
+        # If we keep in the graph the x variable as input of the Apply node,
+        # this can rise the memory usage. That is why the Apply node
+        # take `x_.shape` as input and not `x`.
+        return gof.Apply(self, [x_.shape, values_, ilist_],
+                         [csc_matrix(dtype=x.dtype)])
 
     def perform(self, node, inp, out_):
-        x, values, idx = inp
+        out_shape, values, ilist = inp
         out, = out_
         rows, cols = values.shape
-        assert rows == len(idx)
+        assert rows == len(ilist)
         indptr = numpy.arange(cols + 1) * rows
-        indices = as_strided(idx,
-                             strides=(0, idx.strides[0]),
-                             shape = (cols, idx.shape[0])).flatten()
+        indices = as_strided(ilist,
+                             strides=(0, ilist.strides[0]),
+                             shape=(cols, ilist.shape[0])).flatten()
         data = values.T.flatten()
-        out[0] = scipy.sparse.csc_matrix((data, indices, indptr), shape=x.shape,
-                                    dtype=x.dtype)
+        out[0] = scipy.sparse.csc_matrix((data, indices, indptr),
+                                         shape=out_shape,
+                                         dtype=values.dtype)
 
     def infer_shape(self, node, ishapes):
         x, y, ilist = ishapes
@@ -3356,3 +3397,5 @@ class ConstructSparseFromList(gof.Op):
         gy = theano.tensor.advanced_subtensor1(g_output, *idx_list)
 
         return [gx, gy] + [DisconnectedType()()] * len(idx_list)
+
+construct_sparse_from_list = ConstructSparseFromList()

theano/sparse/tests/test_basic.py

@@ -1027,8 +1027,9 @@ class test_structureddot(unittest.TestCase):
                                      overhead_tol)
 
 
-class DotTests(unittest.TestCase):
+class DotTests(utt.InferShapeTester):
     def setUp(self):
+        super(DotTests, self).setUp()
         x_size = (10, 100)
         y_size = (100, 1000)
         utt.seed_rng()
@@ -1043,48 +1044,47 @@ class DotTests(unittest.TestCase):
             numpy.random.binomial(1, 0.5, y_size), dtype=theano.config.floatX)
         self.y_csc = scipy.sparse.csc_matrix(
             numpy.random.binomial(1, 0.5, y_size), dtype=theano.config.floatX)
+        self.v_10 = numpy.asarray(numpy.random.uniform(-1, 1, 10),
+                                  dtype=theano.config.floatX)
+        self.v_100 = numpy.asarray(numpy.random.uniform(-1, 1, 100),
+                                  dtype=theano.config.floatX)
 
     def test_csr_dense(self):
         x = theano.sparse.csr_matrix('x')
         y = theano.tensor.matrix('y')
+        v = theano.tensor.vector('v')
 
+        for (x, y, x_v, y_v) in [(x, y, self.x_csr, self.y),
+                                 (x, v, self.x_csr, self.v_100),
+                                 (v, x, self.v_10, self.x_csr)]:
             f_a = theano.function([x, y], theano.sparse.dot(x, y))
             f_b = lambda x, y: x * y
 
-        assert _allclose(f_a(self.x_csr, self.y), f_b(self.x_csr, self.y))
+            assert _allclose(f_a(x_v, y_v), f_b(x_v, y_v))
 
             # Test infer_shape
-        f_a = theano.function([x, y], theano.sparse.dot(x, y).shape)
-        f_b = lambda x, y: (x * y).shape
-        assert numpy.all(f_a(self.x_csr, self.y) == f_b(self.x_csr, self.y))
-        topo = f_a.maker.fgraph.toposort()
-        if theano.config.mode != 'FAST_COMPILE':
-            nb = 0
-        else:
-            nb = 1
-        assert sum([isinstance(node.op, (Dot, Usmm, UsmmCscDense))
-                    for node in topo]) == nb
+            self._compile_and_check([x, y], [theano.sparse.dot(x, y)],
+                                    [x_v, y_v],
+                                    (Dot, Usmm, UsmmCscDense))
 
     def test_csc_dense(self):
         x = theano.sparse.csc_matrix('x')
         y = theano.tensor.matrix('y')
+        v = theano.tensor.vector('v')
+
+        for (x, y, x_v, y_v) in [(x, y, self.x_csc, self.y),
+                                 (x, v, self.x_csc, self.v_100),
+                                 (v, x, self.v_10, self.x_csc)]:
 
             f_a = theano.function([x, y], theano.sparse.dot(x, y))
             f_b = lambda x, y: x * y
 
-        assert _allclose(f_a(self.x_csc, self.y), f_b(self.x_csc, self.y))
+            assert _allclose(f_a(x_v, y_v), f_b(x_v, y_v))
 
             # Test infer_shape
-        f_a = theano.function([x, y], theano.sparse.dot(x, y).shape)
-        f_b = lambda x, y: (x * y).shape
-        assert numpy.all(f_a(self.x_csc, self.y) == f_b(self.x_csc, self.y))
-        topo = f_a.maker.fgraph.toposort()
-        if theano.config.mode != 'FAST_COMPILE':
-            nb = 0
-        else:
-            nb = 1
-        assert sum([isinstance(node.op, (Dot, Usmm, UsmmCscDense))
-                    for node in topo]) == nb
+            self._compile_and_check([x, y], [theano.sparse.dot(x, y)],
+                                    [x_v, y_v],
+                                    (Dot, Usmm, UsmmCscDense))
 
     def test_sparse_sparse(self):
         for d1, d2 in [('float32', 'float32'),

theano/tensor/basic.py

@@ -6929,7 +6929,6 @@ class AdvancedSubtensor1(Op):
         out[0] = x.take(i, axis=0, out=o)
 
     def connection_pattern(self, node):
-
         rval = [[True]]
 
         for ipt in node.inputs[1:]:
@@ -6939,17 +6938,18 @@ class AdvancedSubtensor1(Op):
 
     def grad(self, inputs, grads):
         global sparse_module_ref
+        x, ilist = inputs
         gz, = grads
         assert len(inputs) == 2
-        if inputs[0].type.sparse_grad:
+
+        if x.type.sparse_grad:
             if sparse_module_ref is None:
                 import theano.sparse as sparse_module_ref
 
-            rval1 = [sparse_module_ref.ConstructSparseFromList()(
-                                                (inputs[0]), gz, inputs[1])]
+            rval1 = [sparse_module_ref.construct_sparse_from_list(x, gz,
+                                                                  ilist)]
         else:
-            rval1 = [advanced_inc_subtensor1(
-                                        zeros_like(inputs[0]), gz, inputs[1])]
+            rval1 = [advanced_inc_subtensor1(zeros_like(x), gz, ilist)]
         return rval1 + [DisconnectedType()()] * (len(inputs) - 1)
 
     def R_op(self, inputs, eval_points):

theano/tensor/opt.py

@@ -9,6 +9,7 @@ _logger = logging.getLogger('theano.tensor.opt')
 
 import operator
 import itertools
+import StringIO
 import sys
 import traceback
 from itertools import izip
@@ -807,11 +808,15 @@ class ShapeFeature(object):
         else:
             if not isinstance(s, (tuple, list)):
                 raise TypeError('shapes must be tuple/list', (r, s))
+
             if r.ndim != len(s):
+                sio = StringIO.StringIO()
+                theano.printing.debugprint(r, file=sio, print_type=True)
                 raise AssertionError(
                     "Something inferred a shape with %d dimensions "
-                        "for a variable with %d dimensions." % (
-                        len(s), r.ndim))
+                    "for a variable with %d dimensions"
+                    " for the variable:\n%s" % (
+                        len(s), r.ndim, sio.getvalue()))
 
             shape_vars = []
             for i in range(r.ndim):