figuring out the right dtype

theano/scan_module/scan_op.py

@@ -1291,17 +1291,20 @@ class Scan(PureOp):
             if idx >= self.n_mit_mot_outs:
                 Xt_placeholder = Xt.type()
                 Xts.append(Xt_placeholder)
-
-            Xtm1_pos = self.get_input_pos(idx)
-            if Xtm1_pos >= 0:
-                Xtm1 = self_inputs[Xtm1_pos]
-                # It is possible that X[t] is not actually a function of
-                # x[t-1], case in which we can not rely on this information
-                try:
-                    tmp = tensor.grad(Xt.sum(), Xtm1)
-                except ValueError:
-                    tmp = Xt
-                dC_dXt = safe_new(tmp)
+            if Xt not in self.inner_nitsot_outs(self_outputs):
+                dtypes = []
+                states = (self.inner_mitmot(self_inputs) +
+                          self.inner_mitsot(self_inputs) +
+                          self.inner_sitsot(self_inputs))
+
+                for pos, inp in enumerate(states):
+                    if inp in theano.gof.graph.inputs([Xt]):
+                        oidx = self.get_output_pos(pos)
+                        if not isinstance(dC_douts[oidx].type,
+                                          DisconnectedType):
+                            dtypes.append(dC_douts[oidx].dtype)
+                new_dtype = theano.scalar.upcast(*dtypes)
+                dC_dXt = safe_new(Xt, dtype=new_dtype)
             else:
                 if isinstance(dC_douts[idx].type, DisconnectedType):
                     continue
@@ -1331,7 +1334,17 @@ class Scan(PureOp):
                     dC_dinps_t[dx] = tmp
 
         # construct dX_dtm1
-        dC_dXtm1s = [x.type() for x in dC_dinps_t[self.n_seqs:]]
+        dC_dXtm1s = []
+        for pos, x in enumerate(dC_dinps_t[self.n_seqs:]):
+            opos = self.get_output_pos(pos)
+            if opos >= 0:
+                dC_dXtm1s.append(dC_dXts[opos].type())
+                if x.dtype != dC_dXts[opos].dtype:
+                    dC_dinps_t[pos + self.n_seqs] = \
+                            tensor.cast(x,
+                                        dtype=dC_dXts[opos].dtype)
+            else:
+                dC_dXtm1s.append(x.type())
         for dx, dC_dXtm1 in enumerate(dC_dXtm1s):
             dC_dinps_t[dx + self.n_seqs] += dC_dXtm1
         # Construct scan op