Fix mistakes.

theano/sandbox/cuda/tests/test_dnn.py

@@ -16,6 +16,8 @@ import theano.sandbox.cuda.dnn as dnn
 from theano.sandbox.cuda.basic_ops import GpuAllocEmpty, gpu_alloc_empty
 from theano.sandbox.cuda import float32_shared_constructor as shared
 
+from . import test_nnet
+
 # Skip test if cuda_ndarray is not available.
 import theano.sandbox.cuda as cuda
 if not cuda.cuda_available:
@@ -451,10 +453,11 @@ def test_pooling_opt():
                 for n in f.maker.fgraph.toposort()])
 
 
-class test_DnnSoftMax(test_.test_SoftMax):
+class test_DnnSoftMax(test_nnet.test_SoftMax):
     gpu_op = dnn.GpuDnnSoftmax
     gpu_grad_op = dnn.GpuDnnSoftmaxGrad
     mode = mode_with_gpu
+    do_0 = False
     topo_idx = -3
 
     def setUp(self):
@@ -470,7 +473,9 @@ class test_DnnSoftMax(test_.test_SoftMax):
 
         utt.verify_grad(softmax_op, [x_val])
 
-        utt.verify_grad(softmax_op, [x_val2])
+        # Gradient is broken for (n, c, 1, 1) in v3 rc1
+        if cuda.dnn.version() == (2000, 2000):
+            utt.verify_grad(softmax_op, [x_val2])
 
     def test_cudnn_softmax_grad_opt(self):
         # Verify that the SoftmaxGrad -> GpuDnnSoftmaxGrad optimization is

theano/sandbox/cuda/tests/test_nnet.py

@@ -299,7 +299,7 @@ class test_SoftMax(unittest.TestCase):
             check_types
         )
 
-        if do_big:
+        if self.do_big:
             self._cmp(2 << 15, 5, f, f_gpu)
-        if do_0:
+        if self.do_0:
             self._cmp(0, 10, f, f_gpu)