Merge pull request #2480 from nouiz/conv

theano/gof/opt.py

@@ -857,7 +857,9 @@ class LocalMetaOptimizer(LocalOptimizer):
             elif hasattr(input.tag, 'test_value'):
                 givens[input] = theano.shared(
                     input.type.filter(input.tag.test_value),
-                        input.name, borrow=True)
+                    input.name,
+                    broadcastable=input.broadcastable,
+                    borrow=True)
             else:
                 missing.add(input)
         if missing:

theano/sandbox/cuda/blas.py

@@ -1541,6 +1541,18 @@ class GpuConv(GpuOp):
                         to enable them.
         """
         self.border_mode = border_mode
+        if version != -1:
+            raise Exception(
+                """GpuConv with version!=-1 is disabled as we do not
+                test it anymore. It probably work, so you probably can
+                just comment this error and use it. But we want to
+                make sure you know about that. Also, this Op is pretty
+                slow and isn't used by default anymore. We strongly
+                suggest to use GpuCorrMM that is much faster and
+                implement all the functionality (at a cost of some
+                extra memory usage). If you can use cuDNN, that is
+                even better.
+                """)
         self.subsample = subsample
         if logical_img_hw is not None:
             h, w = logical_img_hw

theano/sandbox/cuda/opt.py

@@ -1285,16 +1285,11 @@ def local_conv_gemm(node):
         img, kern = node.inputs
         border_mode = node.op.border_mode
         subsample = node.op.subsample
-        pad = (0,0)
-        if (border_mode == 'full') and (subsample != (1,1)):
-            # need to simulate this via a padded valid convolution
-            pad = 'full'
-            border_mode = 'valid'
-        if (border_mode == 'valid'):
+        if (border_mode == 'valid') or (subsample != (1,1)):
             # need to flip the kernel for valid convolution
             kern = kern[:, :, ::-1, ::-1]
             # By default use GpuCorrMM
-            rval = GpuCorrMM('valid', subsample, pad)(
+            rval = GpuCorrMM(border_mode, subsample)(
                 gpu_contiguous(img), gpu_contiguous(kern))
 
             # call GpuCorrMM_gradWeights if good
@@ -1323,7 +1318,7 @@ def local_conv_gemm(node):
                     # 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('valid', subsample, pad)(
+                        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))
@@ -1331,7 +1326,7 @@ def local_conv_gemm(node):
             # need to dimshuffle the kernel for full convolution
             kern = kern.dimshuffle(1, 0, 2, 3)
             # call GpuCorrMM_gradInputs
-            rval = GpuCorrMM_gradInputs('valid', subsample, pad)(
+            rval = GpuCorrMM_gradInputs('valid', subsample)(
                     gpu_contiguous(kern), gpu_contiguous(img))
         if node.outputs[0].broadcastable != rval.broadcastable:
             # With given shape information, conv2d_fft may return a different
@@ -1413,7 +1408,9 @@ class ConvMetaOptimizer(LocalCudaMetaOptimizer):
 #                                        allow_downcast=True),
                     numpy.require(numpy.random.randn(*shape),
                                   dtype=var.dtype),
-                        var.name, borrow=True)
+                    var.name,
+                    broadcastable=var.broadcastable,
+                    borrow=True)
         # return mapping
         return result
 

theano/sandbox/cuda/tests/test_conv_cuda_ndarray.py

@@ -556,8 +556,8 @@ def _test_subsample(cls, mode, version_valid=[-1], version_full=[-1]):
 
 def test_subsample():
     for t in _test_subsample(None, theano_mode,
-                             version_valid=[-2, -1, 1, 3, 11, 12],
-                             version_full=[-2, -1]):
+                             version_valid=[-1],
+                             version_full=[-1]):
         yield t