Revert the change that would broadcast the output if necessary and just don't…

theano/sandbox/cuda/opt_util.py

@@ -6,7 +6,6 @@ from theano import scalar as scal, Constant
 from theano.gof import local_optimizer
 from theano.tensor import (DimShuffle, get_scalar_constant_value,
                            NotScalarConstantError)
-from theano.tensor.opt import broadcast_like
 
 from theano.sandbox.cuda.basic_ops import (
     GpuFromHost, HostFromGpu, host_from_gpu, GpuDimShuffle, GpuElemwise)
@@ -101,7 +100,9 @@ def output_merge(cls, alpha_in, beta_in, out_in, nd):
                     # other cases are too complex for now
                     return None
                 if W.broadcastable != targ.inputs[out_in].broadcastable:
-                    W = broadcast_like(W, targ.inputs[out_in], node.fgraph)
+                    # May change later to do the broadcast, but it's
+                    # under discussion.
+                    return None
                 inputs = list(targ.inputs)
                 inputs[out_in] = W
                 inputs[beta_in] = _one.clone()

theano/sandbox/cuda/tests/test_dnn.py

@@ -686,15 +686,12 @@ def test_dnn_conv_merge_mouts():
 
 
 def test_dnn_conv_merge_broad():
-    # Make sure that output merge works on broadcasted outputs.
+    # Make sure that we don't apply output_merge on broadcasted values.
     if not cuda.dnn.dnn_available():
         raise SkipTest(cuda.dnn.dnn_available.msg)
     img = T.ftensor4()
     kern = T.ftensor4()
 
-    img_val = numpy.random.random((2, 1, 8, 8)).astype('float32')
-    kern_val = numpy.random.random((3, 1, 4, 5)).astype('float32')
-
     conv = dnn.dnn_conv(img, kern)
 
     lr = numpy.asarray(0.05, dtype='float32')
@@ -702,23 +699,10 @@ def test_dnn_conv_merge_broad():
     # this does broadcasting
     fr = conv + lr
 
-    f1 = theano.function([img, kern], [fr], mode=mode_with_gpu)
-    assert isinstance(f1.maker.fgraph.outputs[0].owner.inputs[0].owner.op,
+    f = theano.function([img, kern], [fr])
+    assert not isinstance(f.maker.fgraph.outputs[0].owner.inputs[0].owner.op,
                           dnn.GpuDnnConv)
 
-    mode = mode_with_gpu
-    mode = mode.excluding('local_dnn_conv_alpha_merge')
-    mode = mode.excluding('local_dnn_conv_output_merge')
-
-    f2 = theano.function([img, kern], [fr], mode=mode)
-    assert not isinstance(f2.maker.fgraph.outputs[0].owner.inputs[0].owner.op,
-                          dnn.GpuDnnConv)
-
-    out_f1 = f1(img_val, kern_val)
-    out_f2 = f2(img_val, kern_val)
-
-    utt.assert_allclose(out_f1, out_f2)
-
 
 def test_dnn_conv_grad():
     if not cuda.dnn.dnn_available() or dnn.version() == -1: