Register the dnn optimizations for AbstractConv.

theano/sandbox/gpuarray/dnn.py

@@ -1258,6 +1258,8 @@ def local_abstractconv_cudnn(node):
                              conv_mode=conv_mode)
     return [rval]
 
+conv_groupopt.register(local_abstactconv_cudnn)
+
 
 @inplace_allocempty(GpuDnnConv, 2)
 def local_dnn_conv_inplace(node, inputs):

theano/sandbox/gpuarray/opt.py

@@ -20,6 +20,7 @@ from theano.tensor.nnet.abstract_conv2d import (BaseAbstractConv2d,
                                                 AbstractConv2d,
                                                 AbstractConv2d_gradWeights,
                                                 AbstractConv2d_gradInputs)
+
 from theano.tests.breakpoint import PdbBreakpoint
 
 from .type import (GpuArrayType, GpuArrayConstant, get_context,
@@ -800,6 +801,10 @@ def local_lift_abstractconv2dgrad(node, context_name):
                     node.inputs[2])]
 
 
+# Register this here so that it goes after the abstract lifting
+register_opt()(conv_groupopt)
+
+
 # This will deal with ops that don't have an explicit transfer but
 # have one of their inputs on the GPU already and the other not on the
 # GPU (to avoid endlessly replacing things).