* fixed csm gradient to work with no kernel map

@TODO: this whole kernel map thing should've probably been implemented as its own op, which would sit between the kernel taps and the CSM op (a "Shuffle" or "TensorView" op). Revisit at a later date

* fixed csm gradient to work with no kernel map
af5445ed · desjagui@atchoum.iro.umontreal.ca · 9121085f · af5445ed
--- a/theano/sparse/basic.py
+++ b/theano/sparse/basic.py
@@ -323,9 +323,12 @@ class CSMGrad(gof.op.Op):
        return gof.Apply(self, [data, gout_data, gout_indices], [g_data])
    def perform(self, node, (data, gout_data, gout_indices), (g_data,)):
-        grad = numpy.zeros_like(data)
+        if self.map is None:
-        grad[self.map] = gout_data
+            g_data[0] = gout_data
-        g_data[0] = grad
+        else:
+            grad = numpy.zeros_like(data)
+            grad[self.map] = gout_data
+            g_data[0] = grad
 csm_grad = CSMGrad
 @gof.local_optimizer([csm_properties])