reorganize code

theano/sandbox/cuda/dnn.py

@@ -13,6 +13,9 @@ from theano.compile.ops import shape_i
 from theano.tensor.nnet import SoftmaxGrad
 from theano.tensor.signal.downsample import (
     DownsampleFactorMax, MaxPoolGrad, AveragePoolGrad)
+from theano.tensor.opt import register_specialize_device
+from theano.sandbox.cuda.type import CudaNdarrayType
+
 from theano.sandbox.cuda import GpuOp
 from theano.sandbox.cuda.basic_ops import (as_cuda_ndarray_variable,
                                            host_from_gpu,
@@ -27,6 +30,12 @@ from theano.sandbox.cuda import gpu_seqopt, register_opt
 
 from theano.sandbox.cuda.nvcc_compiler import NVCC_compiler
 
+from theano.tensor.nnet.abstract_conv2d import (AbstractConv2d,
+                                                AbstractConv2d_gradWeights,
+                                                AbstractConv2d_gradInputs)
+from theano.tensor.opt import register_specialize_device
+
+
 
 def dnn_available():
     if dnn_available.avail is None:
@@ -2439,3 +2448,44 @@ if True:
                 gpu_contiguous(ins[1])
             )
             return [out.dimshuffle(0, 1)]
+
+### AbstractConv Optimizations
+@local_optimizer([AbstractConv2d, AbstractConv2d_gradWeights, AbstractConv2d_gradInputs])
+def local_conv2d_cudnn(node):
+
+    inp1 = node.inputs[0]
+    inp2 = node.inputs[1]
+
+    if not isinstance(inp1.type, CudaNdarrayType) or \
+            not isinstance(inp2.type, CudaNdarrayType):
+        return None
+    if not dnn_available():
+        return None
+
+    if node.op.filters_flip:
+        conv_mode = 'conv'
+    else:
+        conv_mode = 'cross'
+    if (isinstance(node.op, AbstractConv2d)):
+        rval = dnn_conv(inp1, inp2,
+                        border_mode=node.op.border_mode,
+                        subsample=node.op.subsample,
+                        direction_hint='forward',
+                        conv_mode = conv_mode)
+        return [rval]
+    if (isinstance(node.op, AbstractConv2d_gradWeights)):
+        shape = (inp2.shape[1], inp1.shape[1], node.inputs[2][0], node.inputs[2][1])
+        rval = dnn_gradweight(inp1, inp2, shape,
+                              border_mode=node.op.border_mode,
+                              subsample=node.op.subsample,
+                              conv_mode = conv_mode)
+        return [rval]
+    if (isinstance(node.op, AbstractConv2d_gradInputs)):
+        shape = (inp2.shape[0], inp1.shape[1], node.inputs[2][0], node.inputs[2][1])
+        rval = dnn_gradinput(inp1, inp2, shape,
+                             border_mode=node.op.border_mode,
+                             subsample=node.op.subsample,
+                             conv_mode = conv_mode)
+        return [rval]
+register_specialize_device(local_conv2d_cudnn, 'cudnn')
+

theano/sandbox/cuda/opt.py

@@ -75,6 +75,12 @@ from theano.tensor import slinalg
 from theano.tensor.nnet.Conv3D import Conv3D
 from theano.tests.breakpoint import PdbBreakpoint
 
+from theano.tensor.nnet.abstract_conv2d import (BaseAbstractConv2d, AbstractConv2d,
+                                                AbstractConv2d_gradWeights,
+                                                AbstractConv2d_gradInputs)
+from theano.tensor.opt import register_specialize_device
+
+
 try:
     # We need to be able to import this file even if cuda isn't avail.
     from theano.sandbox.cuda import device_properties
@@ -2619,3 +2625,157 @@ optdb.register('local_inplace_gpu_sparse_block_outer',
 
 
 import theano.sandbox.cuda.extra_ops
+
+### Move to Gpu optimization
+@local_optimizer([gpu_from_host,
+                  AbstractConv2d, AbstractConv2d_gradWeights, AbstractConv2d_gradInputs])
+def local_conv2d_gpu_conv(node):
+    """
+    gpu_from_host(AbstractConv) -> AbstractConv(gpu_from_host)
+
+    AbstractConv(host_from_gpu) -> host_from_gpu(AbstractConv)
+    """
+    if isinstance(node.op, GpuFromHost):
+        host_input = node.inputs[0]
+        if host_input.owner and  isinstance(host_input.owner.op, BaseAbstractConv2d):
+
+            conv = host_input.owner.op
+            inps = list(host_input.owner.inputs)
+            inps[0] = as_cuda_ndarray_variable(inps[0])
+            inps[1] = as_cuda_ndarray_variable(inps[1])
+            out = conv(*inps)
+            # out is on the GPU because both inputs are.
+            out = theano.tensor.patternbroadcast(out,
+                                                 node.outputs[0].broadcastable)
+            out.values_eq_approx = values_eq_approx_high_tol
+            return [out]
+
+    if isinstance(node.op, BaseAbstractConv2d):
+        # conv(host_from_gpu) -> host_from_gpu(gpu_conv)
+        inp1 = node.inputs[0]
+        inp2 = node.inputs[1]
+        if ((isinstance(inp1.type, CudaNdarrayType) and
+             isinstance(inp2.type, CudaNdarrayType))):
+            # Both inputs are already directly on the GPU, nothing to do
+            return
+
+        inp1_on_gpu = (isinstance(inp1.type, CudaNdarrayType) or
+                       (inp1.owner and isinstance(inp1.owner.op, HostFromGpu)))
+        inp2_on_gpu = (isinstance(inp2.type, CudaNdarrayType) or
+                       (inp2.owner and isinstance(inp2.owner.op, HostFromGpu)))
+
+        if inp1_on_gpu or inp2_on_gpu:
+            conv = node.op
+            inps = list(node.inputs)
+            inps[0] = as_cuda_ndarray_variable(inps[0])
+            inps[1] = as_cuda_ndarray_variable(inps[1])
+            out = conv(*inps)
+            # out is on the GPU because both inputs are.
+            out = theano.tensor.patternbroadcast(
+                out,
+                node.outputs[0].broadcastable)
+            out.values_eq_approx = values_eq_approx_high_tol
+            # If the original output was on CPU, we have to transfer it
+            if isinstance(node.outputs[0].type, tensor.TensorType):
+                return [tensor.as_tensor_variable(out)]
+            else:
+                return [out]
+register_opt()(local_conv2d_gpu_conv)
+
+
+
+### Corrmm opt
+@local_optimizer([AbstractConv2d])
+def local_conv2d_corrmm(node):
+
+    img, kern = node.inputs
+    if (not isinstance(img.type, CudaNdarrayType) or
+            not isinstance(kern.type, CudaNdarrayType)):
+        return None
+
+    border_mode = node.op.border_mode
+    subsample = node.op.subsample
+    if (border_mode == 'full') and (subsample == (1, 1)):
+        if not node.op.filters_flip:
+            kern = kern[:, :, ::-1, ::-1]
+        # need to dimshuffle the kernel for full convolution
+        kern = kern.dimshuffle(1, 0, 2, 3)
+        # call GpuCorrMM_gradInputs
+        rval = GpuCorrMM_gradInputs('valid', subsample)(
+                gpu_contiguous(kern), gpu_contiguous(img))
+    else:
+        # need to flip the kernel if necessary
+        if node.op.filters_flip:
+            kern = kern[:, :, ::-1, ::-1]
+        # By default use GpuCorrMM
+        rval = GpuCorrMM(border_mode, subsample)(gpu_contiguous(img),
+                                                 gpu_contiguous(kern))
+
+        # call GpuCorrMM_gradWeights if good
+        # (the latter is faster if batchsize * kernelHeight * kernelWidth
+        # is larger than inputChannels * outputHeight * outputWidth.
+        # GpuConv does not always store information on the batchsize and
+        # channels, though, so we only use what information we have.)
+        if ((subsample == (1,1)) and
+            (node.op.imshp is not None) and
+            (None not in node.op.imshp[-2:]) and
+            (node.op.kshp is not None) and
+            (None not in node.op.kshp)):
+            # we know the kernel and output size
+            prod1 = node.op.kshp[0] * node.op.kshp[1]
+            prod2 = ((node.op.imshp[-2] - node.op.kshp[0] + 1) *
+                     (node.op.imshp[-1] - node.op.kshp[1] + 1))
+            if ((node.op.bsize is not None) and
+                    (len(node.op.imshp) == 3) and
+                    (node.op.imshp[0] is not None)):
+                # we also know batchsize and input channels
+                prod1 *= node.op.bsize
+                prod2 *= node.op.imshp[0]
+            # compare to decide
+            if prod1 > prod2:
+                # (we need to wrap the result in as_cuda_ndarray_variable,
+                # because we are not allowed to replace a CudaNdarray with
+                # a DimShuffle instance in a graph optimization)
+                rval = theano.sandbox.cuda.as_cuda_ndarray_variable(
+                    GpuCorrMM_gradWeights(border_mode, subsample)(
+                        gpu_contiguous(img.dimshuffle(1, 0, 2, 3)),
+                        gpu_contiguous(kern.dimshuffle(1, 0, 2, 3))
+                    ).dimshuffle(1, 0, 2, 3))
+    return [rval]
+
+register_specialize_device(local_conv2d_corrmm, 'conv_gemm')
+
+@local_optimizer([AbstractConv2d_gradWeights])
+def local_conv2d_gradweight_corrmm(node):
+
+    img, topgrad, shape = node.inputs
+
+    if not isinstance(img.type, CudaNdarrayType) or \
+            not isinstance(topgrad.type, CudaNdarrayType):
+        return None
+    rval = GpuCorrMM_gradWeights(border_mode=node.op.border_mode,
+                                 subsample=node.op.subsample)(
+        gpu_contiguous(img), gpu_contiguous(topgrad), shape)
+    if node.op.filters_flip:
+        rval = rval[:, :, ::-1, ::-1]
+    rval = tensor.patternbroadcast(rval, node.outputs[0].broadcastable)
+    rval = as_cuda_ndarray_variable(rval)
+    return [rval]
+register_specialize_device(local_conv2d_gradweight_corrmm, 'conv_gemm')
+
+@local_optimizer([AbstractConv2d_gradInputs])
+def local_conv2d_gradinputs_corrmm(node):
+    kern, topgrad, shape = node.inputs
+
+    if not isinstance(kern.type, CudaNdarrayType) or \
+            not isinstance(topgrad.type, CudaNdarrayType):
+        return None
+
+    if node.op.filters_flip:
+        kern = kern[:, :, ::-1, ::-1]
+
+    rval =  GpuCorrMM_gradInputs(border_mode=node.op.border_mode,
+    subsample=node.op.subsample)(
+        gpu_contiguous(kern), gpu_contiguous(topgrad), shape)
+    return [rval]
+register_specialize_device(local_conv2d_gradinputs_corrmm, 'conv_gemm')

theano/sandbox/cuda/tests/test_abstractconv.py

+import unittest
+import numpy
+import copy
+import itertools
+
+import theano
+import theano.tensor as T
+from theano.tests import unittest_tools as utt
+
+from nose.plugins.skip import SkipTest
+import theano.tensor.nnet.conv as conv_ref
+import theano.tensor.nnet.abstract_conv2d as conv
+
+from theano.sandbox.cuda import float32_shared_constructor as gpu_shared
+from theano.compile import shared as cpu_shared
+
+from theano.sandbox.cuda.tests.test_conv_cuda_ndarray import py_conv
+from theano.sandbox.cuda.dnn import dnn_available, dnn_conv, dnn_gradweight, dnn_gradinput
+
+
+if theano.config.mode == 'FAST_COMPILE':
+    mode_with_gpu = theano.compile.mode.get_mode('FAST_RUN').including('gpu')
+    mode_without_gpu = theano.compile.mode.get_mode('FAST_RUN').excluding('gpu')
+else:
+    mode_with_gpu = theano.compile.mode.get_default_mode().including('gpu')
+    mode_without_gpu = theano.compile.get_default_mode().excluding('gpu')
+
+
+class TestConv2d(unittest.TestCase):
+
+    def setUp(self):
+        super(TestConv2d, self).setUp()
+
+        self.inputs_shapes = [(8, 1, 12, 12), (8, 1, 18, 18), (2, 1, 4, 4),
+                              (6, 1, 10, 11), (2, 1, 6, 5), (1, 5, 9, 9)]
+        self.filters_shapes = [(5, 1, 2, 2), (4, 1, 3, 3), (2, 1, 3, 3),
+                               (1, 1, 2, 5), (4, 1, 2, 2), (4, 5, 2, 2)]
+        self.subsamples = [(1, 1), (2, 2), (2, 4)]
+        self.border_modes = ["valid", "full", (0, 0), (1, 1), (5, 5), (5, 2)]
+        self.filters_flip = [True, False]
+
+
+    def get_output_shape(self, inputs_shape, filters_shape, subsample, border_mode):
+        if border_mode == "valid":
+            border_mode = (0, 0)
+        if border_mode == "full":
+            border_mode = (filters_shape[2] - 1, filters_shape[3] - 1)
+        batch_size = inputs_shape[0]
+        num_filters = filters_shape[0]
+        return (batch_size, num_filters,) \
+            + tuple(None if i is None or k is None
+                    else ((i + 2*pad - k) // d + 1)
+                    for i, k, d, pad in zip(inputs_shape[2:], filters_shape[2:],
+                                            subsample, border_mode))
+
+    def run_fwd(self, inputs_shape, filters_shape, ref=dnn_conv,
+                subsample=(1, 1), verify_grad=True, mode=mode_without_gpu,
+                border_mode='valid', filters_flip=True, device='cpu', provide_shape=False):
+
+        inputs_val = numpy.random.random(inputs_shape).astype('float32')
+        filters_val = numpy.random.random(filters_shape).astype('float32')
+        if device == 'gpu':
+            inputs = gpu_shared(inputs_val)
+            filters = gpu_shared(filters_val)
+        else:
+            inputs = theano.tensor.as_tensor_variable(cpu_shared(inputs_val))
+            filters = theano.tensor.as_tensor_variable(cpu_shared(filters_val))
+        if provide_shape:
+            imshp = inputs_shape
+            kshp = filters_shape
+        else:
+            imshp = None
+            kshp = None
+        if filters_flip:
+            conv_mode = 'conv'
+        else:
+            conv_mode = 'cross'
+
+        c_ref = ref(inputs, filters,
+                    border_mode=border_mode,
+                    subsample=subsample,
+                    conv_mode = conv_mode)
+        c = conv.conv2d(inputs, filters,
+                        border_mode=border_mode,
+                        subsample=subsample,
+                        filters_flip=filters_flip,
+                        inputs_shape=imshp,
+                        filters_shape=kshp)
+        f_ref = theano.function([], c_ref, mode=mode)
+        f = theano.function([], c, mode)
+        res_ref = numpy.array(f_ref())
+        res = numpy.array(f())
+        utt.assert_allclose(res_ref, res)
+        if verify_grad:
+            utt.verify_grad(conv.AbstractConv2d(border_mode="valid", imshp=imshp, kshp=kshp,
+                                                bsize=inputs_shape[0], subsample=subsample),
+                            [inputs_val, filters_val],
+                            mode=mode)
+
+    def run_gradweight(self, inputs_shape, filters_shape, output_shape,
+                       ref=dnn_gradweight, subsample=(1, 1), filters_flip=True,
+                       verify_grad=True, mode=mode_without_gpu, border_mode='valid',
+                       device='cpu', provide_shape = False):
+
+        inputs_val = numpy.random.random(inputs_shape).astype('float32')
+        output_val = numpy.random.random(output_shape).astype('float32')
+        if device == 'gpu':
+            inputs = gpu_shared(inputs_val)
+            output = gpu_shared(output_val)
+        else:
+            inputs = theano.tensor.as_tensor_variable(cpu_shared(inputs_val))
+            output = theano.tensor.as_tensor_variable(cpu_shared(output_val))
+        if provide_shape:
+            imshp = inputs_shape
+            kshp = filters_shape
+        else:
+            imshp = None
+            kshp = None
+        if filters_flip:
+            conv_mode = 'conv'
+        else:
+            conv_mode = 'cross'
+        c = conv.AbstractConv2d_gradWeights(border_mode=border_mode,
+                                            filters_flip=filters_flip,
+                                            subsample=subsample,
+                                            imshp = imshp, kshp = kshp)
+        c = c(inputs, output, filters_shape[-2:])
+        c_ref = ref(inputs, output,
+                    filters_shape,
+                    border_mode=border_mode,
+                    subsample=subsample,
+                    conv_mode=conv_mode)
+        f = theano.function([], c, mode)
+        f_ref = theano.function([], c_ref, mode)
+        res_ref = numpy.array(f_ref())
+        res = numpy.array(f())
+        utt.assert_allclose(res_ref, res)
+
+        def abstract_conv2d_gradweight(inputs_val, output_val):
+            conv_op = conv.AbstractConv2d_gradWeights(border_mode=border_mode, subsample=subsample)
+            return conv_op(inputs_val, output_val, filters_shape[-2:])
+
+        if verify_grad:
+            utt.verify_grad(abstract_conv2d_gradweight, [inputs_val, output_val],
+                            mode=mode, eps=1)
+
+
+    def run_gradinput(self, inputs_shape, filters_shape, output_shape, ref=dnn_gradinput,
+                      subsample=(1, 1), filters_flip=True, verify_grad=True, mode=mode_without_gpu,
+                      border_mode='valid', device='cpu', provide_shape = False):
+
+
+        output_val = numpy.random.random(output_shape).astype('float32')
+        filters_val = numpy.random.random(filters_shape).astype('float32')
+        if device == 'gpu':
+            output = gpu_shared(output_val)
+            filters = gpu_shared(filters_val)
+        else:
+            output = theano.tensor.as_tensor_variable(cpu_shared(output_val))
+            filters = theano.tensor.as_tensor_variable(cpu_shared(filters_val))
+        if provide_shape:
+            imshp = inputs_shape
+            kshp = filters_shape
+        else:
+            imshp = None
+            kshp = None
+        if filters_flip:
+            conv_mode = 'conv'
+        else:
+            conv_mode = 'cross'
+        c = conv.AbstractConv2d_gradInputs(border_mode=border_mode,
+                                           subsample=subsample,
+                                           filters_flip=filters_flip,
+                                           imshp=imshp, kshp=kshp)
+        c = c(filters, output, inputs_shape[-2:])
+        c_ref = ref(filters, output, inputs_shape,
+                    border_mode=border_mode, subsample=subsample,
+                    conv_mode=conv_mode)
+        f = theano.function([], c, mode)
+        f_ref = theano.function([], c_ref, mode)
+        res_ref = numpy.array(f_ref())
+        res = numpy.array(f())
+        utt.assert_allclose(res_ref, res)
+
+        def abstract_conv2d_gradinputs(filters_val, output_val):
+            conv_op = conv.AbstractConv2d_gradInputs(border_mode=border_mode, subsample=subsample)
+            return conv_op(filters_val, output_val, inputs_shape[-2:])
+        if verify_grad:
+            utt.verify_grad(abstract_conv2d_gradinputs, [filters_val, output_val],
+                            mode=mode, eps=1)
+
+
+    def test_dnn_conv(self):
+        if not dnn_available():
+            return
+        mode=mode_with_gpu
+        # provide_shape is not used by the CuDNN impementation
+        provide_shape = False
+
+        for (i, f), s, b, flip in itertools.product(
+                zip(self.inputs_shapes, self.filters_shapes),
+                self.subsamples,
+                self.border_modes,
+                self.filters_flip):
+            o = self.get_output_shape(i, f, s, b)
+            self.run_fwd(inputs_shape=i, filters_shape=f, subsample=s,
+                         verify_grad=True, mode=mode, device='gpu',
+                         provide_shape=provide_shape, border_mode=b,
+                         filters_flip=flip)
+            self.run_gradweight(inputs_shape=i, filters_shape=f,
+                                output_shape=o, subsample=s,
+                                verify_grad=True, mode=mode, device='gpu',
+                                provide_shape=provide_shape, border_mode=b,
+                                filters_flip=flip)
+            self.run_gradinput(inputs_shape=i, filters_shape=f,
+                               output_shape=o, subsample=s,
+                               verify_grad=True, mode=mode, device='gpu',
+                               provide_shape=provide_shape, border_mode=b,
+                               filters_flip=flip)
+
+    def test_cormm_conv(self):
+        mode = mode_with_gpu.excluding('cudnn')
+
+        for (i, f), s, b, flip, provide_shape in itertools.product(
+                zip(self.inputs_shapes, self.filters_shapes),
+                self.subsamples,
+                self.border_modes,
+                self.filters_flip,
+                [False, True]):
+
+            o = self.get_output_shape(i, f, s, b)
+            self.run_fwd(inputs_shape=i, filters_shape=f, subsample=s,
+                         verify_grad=True, mode=mode, device='gpu',
+                         provide_shape=provide_shape, border_mode=b,
+                         filters_flip=flip)
+            self.run_gradweight(inputs_shape=i, filters_shape=f,
+                                output_shape=o, subsample=s,
+                                verify_grad=True, mode=mode, device='gpu',
+                                provide_shape=provide_shape, border_mode=b,
+                                filters_flip=flip)
+            self.run_gradinput(inputs_shape=i, filters_shape=f,
+                               output_shape=o, subsample=s,
+                               verify_grad=True, mode=mode, device='gpu',
+                               provide_shape=provide_shape, border_mode=b,
+                               filters_flip=flip)
+
+    def test_cpu_conv(self):
+        mode = mode_without_gpu
+
+        for (i, f), s, b, flip, provide_shape in itertools.product(
+                zip(self.inputs_shapes, self.filters_shapes),
+                self.subsamples,
+                self.border_modes,
+                self.filters_flip,
+                [False, True]):
+
+            o = self.get_output_shape(i, f, s, b)
+            fwd_OK = True
+            gradweight_OK = True
+            gradinput_OK = True
+
+            if not flip:
+                fwd_OK = False
+                gradweight_OK = False
+                gradinput_OK = False
+
+            if b not in ('valid', 'full'):
+                fwd_OK = False
+                gradweight_OK = False
+                gradinput_OK = False
+
+            if (not provide_shape) and (s != (1, 1)) and (b == 'full'):
+                gradweight_OK = False
+                gradinput_OK = False
+
+            if ((s[0] not in (1, 2)) or (s[1] not in (1, 2))) and (b == 'full'):
+                gradweight_OK = False
+                gradinput_OK = False
+
+            if fwd_OK:
+                self.run_fwd(inputs_shape=i, filters_shape=f, subsample=s,
+                             verify_grad=True, mode=mode, device='cpu',
+                             provide_shape=provide_shape, border_mode=b,
+                             filters_flip=flip)
+            else:
+                self.assertRaises(NotImplementedError,
+                                  self.run_fwd,
+                                  inputs_shape=i,
+                                  filters_shape=f,
+                                  subsample=s,
+                                  verify_grad=False,
+                                  mode=mode,
+                                  device='cpu',
+                                  provide_shape=provide_shape,
+                                  border_mode=b,
+                                  filters_flip=flip)
+
+            if gradweight_OK:
+                self.run_gradweight(inputs_shape=i, filters_shape=f,
+                                    output_shape=o, subsample=s,
+                                    verify_grad=False, mode=mode, device='cpu',
+                                    provide_shape=provide_shape, border_mode=b,
+                                    filters_flip=flip)
+            else:
+                self.assertRaises(NotImplementedError,
+                                  self.run_gradweight,
+                                  inputs_shape=i,
+                                  filters_shape=f,
+                                  output_shape=o,
+                                  subsample=s,
+                                  verify_grad=False,
+                                  mode=mode,
+                                  device='cpu',
+                                  provide_shape=provide_shape,
+                                  border_mode=b,
+                                  filters_flip=flip)
+
+            if gradinput_OK:
+                self.run_gradinput(inputs_shape=i, filters_shape=f,
+                                   output_shape=o, subsample=s,
+                                   verify_grad=False, mode=mode, device='cpu',
+                                   provide_shape=provide_shape, border_mode=b,
+                                   filters_flip=flip)
+            else:
+                self.assertRaises(NotImplementedError,
+                                  self.run_gradinput,
+                                  inputs_shape=i,
+                                  filters_shape=f,
+                                  output_shape=o,
+                                  subsample=s,
+                                  verify_grad=False,
+                                  mode=mode,
+                                  device='cpu',
+                                  provide_shape=provide_shape,
+                                  border_mode=b,
+                                  filters_flip=flip)

theano/tensor/nnet/abstract_conv2d.py

@@ -15,20 +15,8 @@ from theano.tensor import TensorType
 from theano.gof import Apply, Op
 from theano.gof import local_optimizer
 
-from theano.sandbox.cuda import register_opt as register_gpu
 from theano.tensor.opt import register_specialize_device
-
-
-### Gpu related optimization (to be moved in sandbox/cuda)
-from theano.sandbox.cuda.basic_ops import (
-    as_cuda_ndarray_variable,
-    gpu_contiguous, gpu_from_host, host_from_gpu,
-    GpuFromHost, HostFromGpu
-    )
 from theano.sandbox.cuda.type import CudaNdarrayType
-from theano.sandbox.cuda.dnn import dnn_available, dnn_conv, dnn_gradweight, dnn_gradinput
-from theano.sandbox.cuda.blas import GpuCorrMM, GpuCorrMM_gradWeights, GpuCorrMM_gradInputs
-from theano.sandbox.cuda.opt import values_eq_approx_high_tol
 
 
 ## Cpu implementation
@@ -36,6 +24,7 @@ from theano.tensor.nnet import conv2d as cpu_conv2d, ConvOp
 from theano.tensor.nnet.ConvGrad3D import convGrad3D
 from theano.tensor.nnet.ConvTransp3D import convTransp3D
 
+
 _logger = logging.getLogger("theano.tensor.nnet.conv2d")
 
 
@@ -330,202 +319,6 @@ class AbstractConv2d_gradInputs(BaseAbstractConv2d):
     def connection_pattern(self, node):
         return [[1], [1], [0]]  # no connection to height, width
 
-
-### Move to Gpu optimization
-@local_optimizer([gpu_from_host,
-                  AbstractConv2d, AbstractConv2d_gradWeights, AbstractConv2d_gradInputs])
-def local_conv2d_gpu_conv(node):
-    """
-    gpu_from_host(AbstractConv) -> AbstractConv(gpu_from_host)
-
-    AbstractConv(host_from_gpu) -> host_from_gpu(AbstractConv)
-    """
-    if isinstance(node.op, GpuFromHost):
-        host_input = node.inputs[0]
-        if host_input.owner and  isinstance(host_input.owner.op, BaseAbstractConv2d):
-
-            conv = host_input.owner.op
-            inps = list(host_input.owner.inputs)
-            inps[0] = as_cuda_ndarray_variable(inps[0])
-            inps[1] = as_cuda_ndarray_variable(inps[1])
-            out = conv(*inps)
-            # out is on the GPU because both inputs are.
-            out = theano.tensor.patternbroadcast(out,
-                                                 node.outputs[0].broadcastable)
-            out.values_eq_approx = values_eq_approx_high_tol
-            return [out]
-
-    if isinstance(node.op, BaseAbstractConv2d):
-        # conv(host_from_gpu) -> host_from_gpu(gpu_conv)
-        inp1 = node.inputs[0]
-        inp2 = node.inputs[1]
-        if ((isinstance(inp1.type, CudaNdarrayType) and
-             isinstance(inp2.type, CudaNdarrayType))):
-            # Both inputs are already directly on the GPU, nothing to do
-            return
-
-        inp1_on_gpu = (isinstance(inp1.type, CudaNdarrayType) or
-                       (inp1.owner and isinstance(inp1.owner.op, HostFromGpu)))
-        inp2_on_gpu = (isinstance(inp2.type, CudaNdarrayType) or
-                       (inp2.owner and isinstance(inp2.owner.op, HostFromGpu)))
-
-        if inp1_on_gpu or inp2_on_gpu:
-            conv = node.op
-            inps = list(node.inputs)
-            inps[0] = as_cuda_ndarray_variable(inps[0])
-            inps[1] = as_cuda_ndarray_variable(inps[1])
-            out = conv(*inps)
-            # out is on the GPU because both inputs are.
-            out = theano.tensor.patternbroadcast(
-                out,
-                node.outputs[0].broadcastable)
-            out.values_eq_approx = values_eq_approx_high_tol
-            # If the original output was on CPU, we have to transfer it
-            if isinstance(node.outputs[0].type, TensorType):
-                return [as_tensor_variable(out)]
-            else:
-                return [out]
-register_gpu()(local_conv2d_gpu_conv)
-
-
-### Cudnn Opt
-@local_optimizer([AbstractConv2d, AbstractConv2d_gradWeights, AbstractConv2d_gradInputs])
-def local_conv2d_cudnn(node):
-
-    inp1 = node.inputs[0]
-    inp2 = node.inputs[1]
-
-    if not isinstance(inp1.type, CudaNdarrayType) or \
-            not isinstance(inp2.type, CudaNdarrayType):
-        return None
-    if not dnn_available():
-        return None
-
-    if node.op.filters_flip:
-        conv_mode = 'conv'
-    else:
-        conv_mode = 'cross'
-    if (isinstance(node.op, AbstractConv2d)):
-        rval = dnn_conv(inp1, inp2,
-                        border_mode=node.op.border_mode,
-                        subsample=node.op.subsample,
-                        direction_hint='forward',
-                        conv_mode = conv_mode)
-        return [rval]
-    if (isinstance(node.op, AbstractConv2d_gradWeights)):
-        shape = (inp2.shape[1], inp1.shape[1], node.inputs[2][0], node.inputs[2][1])
-        rval = dnn_gradweight(inp1, inp2, shape,
-                              border_mode=node.op.border_mode,
-                              subsample=node.op.subsample,
-                              conv_mode = conv_mode)
-        return [rval]
-    if (isinstance(node.op, AbstractConv2d_gradInputs)):
-        shape = (inp2.shape[0], inp1.shape[1], node.inputs[2][0], node.inputs[2][1])
-        rval = dnn_gradinput(inp1, inp2, shape,
-                             border_mode=node.op.border_mode,
-                             subsample=node.op.subsample,
-                             conv_mode = conv_mode)
-        return [rval]
-register_specialize_device(local_conv2d_cudnn, 'cudnn')
-
-### Corrmm opt
-@local_optimizer([AbstractConv2d])
-def local_conv2d_corrmm(node):
-
-    img, kern = node.inputs
-    if (not isinstance(img.type, CudaNdarrayType) or
-            not isinstance(kern.type, CudaNdarrayType)):
-        return None
-
-    border_mode = node.op.border_mode
-    subsample = node.op.subsample
-    if (border_mode == 'full') and (subsample == (1, 1)):
-        if not node.op.filters_flip:
-            kern = kern[:, :, ::-1, ::-1]
-        # need to dimshuffle the kernel for full convolution
-        kern = kern.dimshuffle(1, 0, 2, 3)
-        # call GpuCorrMM_gradInputs
-        rval = GpuCorrMM_gradInputs('valid', subsample)(
-                gpu_contiguous(kern), gpu_contiguous(img))
-    else:
-        # need to flip the kernel if necessary
-        if node.op.filters_flip:
-            kern = kern[:, :, ::-1, ::-1]
-        # By default use GpuCorrMM
-        rval = GpuCorrMM(border_mode, subsample)(gpu_contiguous(img),
-                                                 gpu_contiguous(kern))
-
-        # call GpuCorrMM_gradWeights if good
-        # (the latter is faster if batchsize * kernelHeight * kernelWidth
-        # is larger than inputChannels * outputHeight * outputWidth.
-        # GpuConv does not always store information on the batchsize and
-        # channels, though, so we only use what information we have.)
-        if ((subsample == (1,1)) and
-            (node.op.imshp is not None) and
-            (None not in node.op.imshp[-2:]) and
-            (node.op.kshp is not None) and
-            (None not in node.op.kshp)):
-            # we know the kernel and output size
-            prod1 = node.op.kshp[0] * node.op.kshp[1]
-            prod2 = ((node.op.imshp[-2] - node.op.kshp[0] + 1) *
-                     (node.op.imshp[-1] - node.op.kshp[1] + 1))
-            if ((node.op.bsize is not None) and
-                    (len(node.op.imshp) == 3) and
-                    (node.op.imshp[0] is not None)):
-                # we also know batchsize and input channels
-                prod1 *= node.op.bsize
-                prod2 *= node.op.imshp[0]
-            # compare to decide
-            if prod1 > prod2:
-                # (we need to wrap the result in as_cuda_ndarray_variable,
-                # because we are not allowed to replace a CudaNdarray with
-                # a DimShuffle instance in a graph optimization)
-                rval = theano.sandbox.cuda.as_cuda_ndarray_variable(
-                    GpuCorrMM_gradWeights(border_mode, subsample)(
-                        gpu_contiguous(img.dimshuffle(1, 0, 2, 3)),
-                        gpu_contiguous(kern.dimshuffle(1, 0, 2, 3))
-                    ).dimshuffle(1, 0, 2, 3))
-    return [rval]
-
-register_specialize_device(local_conv2d_corrmm, 'conv_gemm')
-
-@local_optimizer([AbstractConv2d_gradWeights])
-def local_conv2d_gradweight_corrmm(node):
-
-    img, topgrad, shape = node.inputs
-
-    if not isinstance(img.type, CudaNdarrayType) or \
-            not isinstance(topgrad.type, CudaNdarrayType):
-        return None
-    rval = GpuCorrMM_gradWeights(border_mode=node.op.border_mode,
-                                 subsample=node.op.subsample)(
-        gpu_contiguous(img), gpu_contiguous(topgrad), shape)
-    if node.op.filters_flip:
-        rval = rval[:, :, ::-1, ::-1]
-    rval = patternbroadcast(rval, node.outputs[0].broadcastable)
-    rval = as_cuda_ndarray_variable(rval)
-    return [rval]
-register_specialize_device(local_conv2d_gradweight_corrmm, 'conv_gemm')
-
-@local_optimizer([AbstractConv2d_gradInputs])
-def local_conv2d_gradinputs_corrmm(node):
-    kern, topgrad, shape = node.inputs
-
-    if not isinstance(kern.type, CudaNdarrayType) or \
-            not isinstance(topgrad.type, CudaNdarrayType):
-        return None
-
-    if node.op.filters_flip:
-        kern = kern[:, :, ::-1, ::-1]
-
-    rval =  GpuCorrMM_gradInputs(border_mode=node.op.border_mode,
-    subsample=node.op.subsample)(
-        gpu_contiguous(kern), gpu_contiguous(topgrad), shape)
-    return [rval]
-register_specialize_device(local_conv2d_gradinputs_corrmm, 'conv_gemm')
-
-
-
 ### Cpu Optmization
 @local_optimizer([AbstractConv2d])
 def local_conv2d_cpu(node):