Merge pull request #2276 from f0k/dnn-grad-grads

theano/sandbox/cuda/dnn.py

@@ -449,6 +449,21 @@ class GpuDnnConvGradW(GpuDnnConvBase):
     path_flag = 'CUDNN_CONVOLUTION_WEIGHT_GRAD'
     conv_op = 'cudnnConvolutionBackwardFilter'
 
+    def grad(self, inp, grads):
+        img, top, desc = inp
+        kerns, = grads
+
+        kerns = gpu_contiguous(kerns)
+
+        d_img = GpuDnnConvGradI()(kerns, top, desc)
+        d_top = GpuDnnConv()(img, kerns, desc)
+
+        return d_img, d_top, theano.gradient.DisconnectedType()()
+
+    def connection_pattern(self, node):
+        # not connected to desc
+        return [[1], [1], [0]]
+
 
 class GpuDnnConvGradI(GpuDnnConvBase):
     """
@@ -466,6 +481,21 @@ class GpuDnnConvGradI(GpuDnnConvBase):
     path_flag = 'CUDNN_CONVOLUTION_DATA_GRAD'
     conv_op = 'cudnnConvolutionBackwardData'
 
+    def grad(self, inp, grads):
+        kerns, top, desc = inp
+        img, = grads
+
+        img = gpu_contiguous(img)
+
+        d_kerns = GpuDnnConvGradW()(img, top, desc)
+        d_top = GpuDnnConv()(img, kerns, desc)
+
+        return d_kerns, d_top, theano.gradient.DisconnectedType()()
+
+    def connection_pattern(self, node):
+        # not connected to desc
+        return [[1], [1], [0]]
+
 
 def dnn_conv(img, kerns, border_mode='valid', subsample=(1, 1),
              conv_mode='conv', direction_hint=None):

theano/sandbox/cuda/tests/test_conv_cuda_ndarray.py

@@ -1005,7 +1005,7 @@ def conv_grad(mode, bs, ch, nf, rImg1, rImg2, rFlt1, rFlt2, subsample, op):
         # skip if the reference implementation can't do it
         pass
 
-    f = theano.function([i, k], outputs, mode=theano_mode)
+    f = theano.function([i, k], outputs, mode=theano_mode.excluding('conv_dnn', 'conv_gemm'))
 
     allvals = f(npy_img, npy_kern)