Merge pull request #1178 from pascanur/nicolasDLT_scan_bug

theano/scan_module/scan_op.py

@@ -1344,8 +1344,8 @@ class Scan(PureOp):
                     j_inp_idx = self.get_input_pos(jidx) + 1
                     if connection_pattern[j_inp_idx][iidx] == True:
                         for k in xrange(len(connection_pattern)):
-                            if connection_pattern[k][iidx]:
-                                connection_pattern[k][jidx] = True
+                            if connection_pattern[k][jidx]:
+                                connection_pattern[k][iidx] = True
         return connection_pattern
 
     ### GRAD FUNCTION

theano/scan_module/scan_opt.py

@@ -1463,10 +1463,10 @@ class PushOutDot1(gof.Optimizer):
                     inp2 = x.owner.inputs[1]
 
                     if inp1 in seqs or inp2 in seqs:
-                        new_scan_out = inp2
-
-                        if inp2 in seqs:
                         new_scan_out = inp1
+
+                        if inp1 in seqs:
+                            new_scan_out = inp2
                         idx = sitsot_outs.index(out)
                         # We've found our pattern and need to construct a new
                         # scan node to replace this one. For this we need to
@@ -1535,6 +1535,8 @@ class PushOutDot1(gof.Optimizer):
                                         outer_non_seqs)
 
                         new_outs = new_op(*_scan_inputs)
+                        if type(new_outs) not in (list, tuple):
+                            new_outs = [new_outs]
 
                         # We need now to pair correctly the new outputs with the
                         # old ones

theano/scan_module/tests/test_scan.py

@@ -3310,6 +3310,32 @@ class T_Scan(unittest.TestCase):
         # disconnected gradient a non disconnected type was returned
         tensor.grad((m * m2).sum(), v)
 
+    def test_disconnected_gradient2(self):
+        v = tensor.vector('v')
+        m = tensor.matrix('m')
+        u0 = tensor.zeros((7,))
+
+        [u, m2], _ = theano.scan(lambda x, u: [x+u, u+v],
+                                 sequences=m,
+                                 outputs_info=[u0, None])
+        # This used to raise an exception with older versions becasue
+        # scan could not detect the connection between `m2` and `x`
+        tensor.grad(m2.sum(), m)
+
+    def test_dot_optimization(self):
+        A = tensor.matrix('A')
+        B = tensor.matrix('B')
+        S, _ = theano.scan(lambda x1,x2, u: u + tensor.dot(x1,x2),
+                           sequences = [A.dimshuffle(0, 1, 'x'),
+                                        B.dimshuffle(0,'x', 1)],
+                           outputs_info=[tensor.zeros_like(A)])
+        f = theano.function([A,B], S.owner.inputs[0][-1])
+        rng = numpy.random.RandomState(utt.fetch_seed())
+        vA = rng.uniform(size=(5,5))
+        vB = rng.uniform(size=(5,5))
+        assert numpy.allclose(f(vA, vB), numpy.dot(vA.T, vB))
+
+
     def test_pregreedy_optimizer(self):
         W = tensor.zeros((5, 4))
         bv = tensor.zeros((5,))