replaced zeros_like(x) with x.zeros_like() in order to support sparse

theano/gradient.py

@@ -512,7 +512,7 @@ def grad(cost, wrt, g_cost = None, consider_constant = None, warn_type = 'ignore
 
     #the gradient of the constants is 0
     for const in consider_constant:
-        grad_dict[const] = tensor.zeros_like(const)
+        grad_dict[const] = const.zeros_like()
 
     #variables that do not influence the cost have zero gradient.
     #if wrt is such a variable, populate the grad_dict with this info
@@ -552,7 +552,7 @@ def grad(cost, wrt, g_cost = None, consider_constant = None, warn_type = 'ignore
             term_dict[node] = list(input_grads)
             for i in xrange(len(term_dict[node])):
                 if term_dict[node][i] is None:
-                    term_dict[node][i] = tensor.zeros_like(node.inputs[i])
+                    term_dict[node][i] = node.inputs[i].zeros_like()
         return term_dict[node]
 
 
@@ -586,7 +586,7 @@ def grad(cost, wrt, g_cost = None, consider_constant = None, warn_type = 'ignore
             else:
                 #this variable is not connected to the cost in the computational
                 #graph so the gradient on it is zero
-                grad_dict[var] = tensor.zeros_like(var)
+                grad_dict[var] = var.zeros_like()
         return grad_dict[var]
 
 
@@ -662,7 +662,7 @@ def grad_sources_inputs(sources, graph_inputs, warn_type = 'ignored'):
                     [access_grad_cache(var) for var in node.outputs]))
             for i in xrange(len(term_dict[node])):
                 if term_dict[node][i] is None:
-                    term_dict[node][i] = tensor.zeros_like(node.inputs[i])
+                    term_dict[node][i] = node.inputs[i].zeros_like()
         return term_dict[node]
 
 
@@ -694,7 +694,7 @@ def grad_sources_inputs(sources, graph_inputs, warn_type = 'ignored'):
             else:
                 #this variable is not connected to the cost in the computational
                 #graph so the gradient on it is zero
-                grad_dict[var] = tensor.zeros_like(var)
+                grad_dict[var] = var.zeros_like()
         return grad_dict[var]