Changed graph traversal routines in scan_utils.py to not visit any subgraph more than once

theano/scan_module/scan_utils.py

@@ -105,7 +105,7 @@ class until(object):
         assert self.condition.ndim == 0
 
 
-def traverse(out, x, x_copy, d):
+def traverse(out, x, x_copy, d, visited=None):
     ''' Function used by scan to parse the tree and figure out which nodes
     it needs to replace. There are two options :
         1) x and x_copy or on host, then you would replace x with x_copy
@@ -114,6 +114,15 @@ def traverse(out, x, x_copy, d):
     This happens because initially shared variables are on GPU .. which is
     fine for the main computational graph but confuses things a bit for the
     inner graph of scan '''
+    # ``visited`` is a set of nodes that are already known and do not need
+    # to be checked again, speeding up the traversal of interconnected graphs.
+    # if a ``visited`` set is given, it will be updated in-place so the callee
+    # knows which nodes we have seen.
+    if visited is None:
+        visited = set()
+    if out in visited:
+        return d
+    visited.add(out)
     import theano.sandbox.cuda as cuda
     if out == x:
         d[out] = cuda.gpu_from_host(x_copy)
@@ -127,7 +136,7 @@ def traverse(out, x, x_copy, d):
         return d
     else:
         for inp in out.owner.inputs:
-            d = traverse(inp, x, x_copy, d)
+            d = traverse(inp, x, x_copy, d, visited)
         return d
 
 
@@ -988,7 +997,13 @@ def forced_replace(out, x, y):
     if out is None:
         return None
 
-    def traverse(graph, x):
+    # ``visited`` is a set of nodes that are already known and do not need
+    # to be checked again, speeding up the traversal of interconnected graphs.
+    visited = set()
+    def local_traverse(graph, x):
+        if graph in visited:
+            return []
+        visited.add(graph)
         if equal_computations([graph], [x]):
             return [graph]
         elif not graph.owner:
@@ -996,7 +1011,7 @@ def forced_replace(out, x, y):
         else:
             rval = []
             for inp in graph.owner.inputs:
-                rval += traverse(inp, x)
+                rval += local_traverse(inp, x)
             return rval
-    to_replace = traverse(out, x)
+    to_replace = local_traverse(out, x)
     return clone(out, replace=OrderedDict((v, y) for v in to_replace))