Don't enable the GPU version of conv3d ops, as they aren't tested and bugged and VERY slow!

theano/sandbox/cuda/GpuConv3D.py

@@ -7,7 +7,7 @@ from theano.sandbox.cuda.basic_ops import (as_cuda_ndarray_variable,
                                            host_from_gpu, HostFromGpu)
 from theano.misc import strutil
 from theano.tensor.nnet.Conv3D import Conv3D
-from theano.sandbox.cuda.opt import register_opt
+from theano.sandbox.cuda.opt import gpu_optimizer
 from theano.sandbox.cuda import CudaNdarrayType, GpuOp
 
 
@@ -288,7 +288,6 @@ conv_rows_stack( float* img, float* kern, float* bias, float* out,
 gpu_convd = GpuConv3D()
 
 
-@register_opt()
 @local_optimizer([Conv3D])
 def local_gpu_conv3d(node):
     if isinstance(node.op, Conv3D):
@@ -300,3 +299,5 @@ def local_gpu_conv3d(node):
                                                 as_cuda_ndarray_variable(W),
                                                 as_cuda_ndarray_variable(b),
                                                 d))]
+# Not enabled by default as we don't want people to use it.
+gpu_optimizer.register("local_gpu_conv3d", local_gpu_conv3d)

theano/sandbox/cuda/GpuConvGrad3D.py

@@ -7,7 +7,7 @@ from theano.sandbox.cuda.basic_ops import as_cuda_ndarray_variable
 from theano.misc import strutil
 
 from theano.tensor.nnet.ConvGrad3D import ConvGrad3D
-from theano.sandbox.cuda.opt import register_opt
+from theano.sandbox.cuda.opt import gpu_optimizer
 from theano.sandbox.cuda import (CudaNdarrayType, HostFromGpu,
                                  host_from_gpu, GpuOp)
 
@@ -341,9 +341,8 @@ convgrad_rows_stack( float* img, float* dCdH, float* dCdW,
 gpu_conv_grad3d = GpuConvGrad3D()
 
 
-@register_opt()
 @local_optimizer([ConvGrad3D])
-def local_gpu_conv_gradd(node):
+def local_gpu_conv_grad3d(node):
     if isinstance(node.op, ConvGrad3D):
         if numpy.any([i.owner and isinstance(i.owner.op, HostFromGpu)
                       for i in node.inputs]):
@@ -354,3 +353,5 @@ def local_gpu_conv_gradd(node):
                     d,
                     WShape,
                     as_cuda_ndarray_variable(dCdH)))]
+# Not enabled by default as we don't want people to use it.
+gpu_optimizer.register("local_gpu_conv_grad3d", local_gpu_conv_grad3d)

theano/sandbox/cuda/GpuConvTransp3D.py

@@ -8,7 +8,7 @@ from theano.tensor.nnet.ConvTransp3D import ConvTransp3D
 from theano.gof import local_optimizer
 
 from theano.sandbox.cuda.basic_ops import as_cuda_ndarray_variable
-from theano.sandbox.cuda.opt import register_opt
+from theano.sandbox.cuda.opt import gpu_optimizer
 from theano.sandbox.cuda import (CudaNdarrayType, HostFromGpu,
                                  host_from_gpu, GpuOp)
 
@@ -347,15 +347,16 @@ conv_transp_rows_stack( float* H, float* kern, float* bias, float* R,
 gpu_conv_transpd = GpuConvTransp3D()
 
 
-@register_opt()
 @local_optimizer([ConvTransp3D])
-def local_gpu_conv_transpd(node):
+def local_gpu_conv_transp3d(node):
     if isinstance(node.op, ConvTransp3D):
         if numpy.any([i.owner and isinstance(i.owner.op, HostFromGpu)
                       for i in node.inputs]):
             if numpy.all([o.type.dtype == 'float32' for o in node.outputs]):
                 W, b, d, H, RShape = node.inputs
                 return [host_from_gpu(gpu_conv_transpd(W, b, d, H, RShape))]
+# Not enabled by default as we don't want people to use it.
+gpu_optimizer.register("local_gpu_conv_transp3d", local_gpu_conv_transp3d)
 
 
 #If the input size wasn't a multiple of D we may need to cause some automatic padding to get the right size of reconstruction