Merge pull request #1614 from pascanur/new_fix_grad_scan

theano/scan_module/scan_op.py

@@ -1561,19 +1561,16 @@ class Scan(PureOp):
         for idx in xrange(self.n_mit_mot + self.n_mit_sot):
             mintap = numpy.min(self.tap_array[idx])
             maxtap = numpy.max(self.tap_array[idx])
+            if idx < self.n_mit_mot:
+                outmaxtap = numpy.max(self.mitmot_out_taps()[idx])
+            else:
+                outmaxtap = 0
             seq = outs[idx]
             for k in self.tap_array[idx]:
-                if maxtap < 0:
-                    dim_offset = abs(maxtap)
-                else:
-                    dim_offset = 0
-                if maxtap == mintap and maxtap != 0:
-                    nw_seq = seq[:abs(maxtap)]
-                elif maxtap - k != 0:
-                    nw_seq = seq[dim_offset + k - mintap - 1:\
-                                 -(maxtap - k + 1)][::-1]
+                if outmaxtap -k != 0:
+                    nw_seq = seq[k - mintap: -(outmaxtap-k)][::-1]
                 else:
-                    nw_seq = seq[dim_offset + k - mintap - 1: -1][::-1]
+                    nw_seq = seq[k - mintap:][::-1]
                 outer_inp_seqs.append(nw_seq)
         outer_inp_seqs += [
             x[:-1][::-1] for x in self.outer_sitsot_outs(outs)]
@@ -1627,7 +1624,11 @@ class Scan(PureOp):
         n_mitmot_inps = 0
 
         for idx in xrange(self.n_mit_mot):
-            outer_inp_mitmot.append(dC_douts[idx][::-1])
+            if isinstance(dC_douts[idx].type, DisconnectedType):
+                out = outs[idx]
+                outer_inp_mitmot.append(tensor.zeros_like(out))
+            else:
+                outer_inp_mitmot.append(dC_douts[idx][::-1])
             mitmot_inp_taps.append([])
             mitmot_out_taps.append([])
             undefined = False
@@ -1648,7 +1649,7 @@ class Scan(PureOp):
                     if _sh in gof.graph.inputs([dC_dinps_t[ins_pos]]):
                         undefined = True
 
-                n_mitmot_inps_ += 1
+                n_mitmot_inps += 1
                 ins_pos += 1
                 n_mitmot_outs += 1
                 mitmot_inp_taps[idx].append(-self.tap_array[idx][jdx])

theano/scan_module/tests/test_scan.py

@@ -3577,10 +3577,8 @@ class T_Scan(unittest.TestCase):
         assert not opt_obj.belongs_to_set(scan_node2, [scan_node1])
 
     def test_remove_constants_and_unused_inputs_scan_non_seqs(self):
-        """Test the opt remove_constants_and_unused_inputs_scan for
-        non sequences.
-
-        """
+        #Test the opt remove_constants_and_unused_inputs_scan for
+        #non sequences.
         W = theano.tensor.matrix(name='W')
         v = theano.tensor.ivector(name='v')
         y1, _ = theano.scan(lambda i, W: W[i], sequences=v,
@@ -3616,10 +3614,7 @@ class T_Scan(unittest.TestCase):
             assert (len(inp) == len(set(inp)))
 
     def test_remove_constants_and_unused_inputs_scan_seqs(self):
-        """
-        Test the opt remove_constants_and_unused_inputs_scan for sequences.
-
-        """
+        #Test the opt remove_constants_and_unused_inputs_scan for sequences.
         W = theano.tensor.matrix(name='W')
         v = theano.tensor.ivector(name='v')
         vv = theano.tensor.matrix(name='vv')
@@ -3661,6 +3656,39 @@ class T_Scan(unittest.TestCase):
             inp = scan_node.op.outer_non_seqs(scan_node)
             assert len(inp) == 1
 
+    def test_hessian_bug_grad_grad_two_scans(self):
+        #Bug reported by Bitton Tenessi
+
+        W_flat = tensor.fvector(name='W')
+        W_flat.tag.test_value=numpy.ones((8,), dtype=numpy.float32)
+        W = W_flat.reshape((2,2,2))
+
+        def loss_outer(i_outer, sum_outer, W):
+
+            def loss_inner(i_inner, sum_inner, W):
+
+                return sum_inner + (W**2).sum().sum().sum()
+
+            result_inner, _ = theano.scan(fn=loss_inner,
+               outputs_info = tensor.as_tensor_variable(
+                   numpy.asarray(0, dtype=numpy.float32)),
+               sequences=tensor.arange(1, dtype='int32'),
+               non_sequences=[W],
+               )
+            return sum_outer + result_inner[-1]
+
+        result_outer, _ = theano.scan(fn=loss_outer,
+               outputs_info = tensor.as_tensor_variable(
+                   numpy.asarray(0, dtype=numpy.float32)),
+               sequences=tensor.arange(1, dtype='int32'),
+               non_sequences=[W],
+               )
+
+        cost = result_outer[-1]
+        H = theano.gradient.hessian(cost, W_flat)
+        f = theano.function([W_flat], H)
+        f(numpy.ones((8,), dtype='float32'))
+
 
 def test_speed():
     #