diff --git a/paddle/fluid/operators/softmax_op.h b/paddle/fluid/operators/softmax_op.h
index 34c35ee421d9adafed758387b973e804760c5a2e..1205bd0587f32caae04c27ecea581fc17988507f 100644
--- a/paddle/fluid/operators/softmax_op.h
+++ b/paddle/fluid/operators/softmax_op.h
@@ -35,8 +35,8 @@ class SoftmaxKernel : public framework::OpKernel<T> {
     auto flattened_dims = framework::flatten_to_2d(dims, dims.size() - 1);
     framework::LoDTensor flattened_x;
     framework::LoDTensor flattened_out;
-    flattened_x.ShareDataWith(*X);
-    flattened_out.ShareDataWith(*Out);
+    flattened_x.ShareDataWith(*X).Resize(flattened_dims);
+    flattened_out.ShareDataWith(*Out).Resize(flattened_dims);
 
     math::SoftmaxFunctor<DeviceContext, T>()(
         context.template device_context<DeviceContext>(), &flattened_x,
@@ -60,9 +60,9 @@ class SoftmaxGradKernel : public framework::OpKernel<T> {
     framework::LoDTensor flattened_out;
     framework::LoDTensor flattened_d_out;
     framework::LoDTensor flattened_d_x;
-    flattened_out.ShareDataWith(*Out);
-    flattened_d_out.ShareDataWith(*dOut);
-    flattened_d_x.ShareDataWith(*dX);
+    flattened_out.ShareDataWith(*Out).Resize(flattened_dims);
+    flattened_d_out.ShareDataWith(*dOut).Resize(flattened_dims);
+    flattened_d_x.ShareDataWith(*dX).Resize(flattened_dims);
 
     math::SoftmaxGradFunctor<DeviceContext, T>()(
         context.template device_context<DeviceContext>(), &flattened_out,
diff --git a/python/paddle/fluid/tests/unittests/test_softmax_op.py b/python/paddle/fluid/tests/unittests/test_softmax_op.py
index 0ab581cfb0ea0ff2205450b8e62edb8bf3c51707..70ad05597c4a160cf6a25aeb3c379320cef69c63 100644
--- a/python/paddle/fluid/tests/unittests/test_softmax_op.py
+++ b/python/paddle/fluid/tests/unittests/test_softmax_op.py
@@ -26,15 +26,22 @@ def stable_softmax(x):
 
 
 class TestSoftmaxOp(OpTest):
+    def get_x_shape(self):
+        return [10, 10]
+
     def setUp(self):
         self.op_type = "softmax"
         self.use_cudnn = False
         self.use_mkldnn = False
         self.dtype = np.float32
         self.init_kernel_type()
+        self.shape = self.get_x_shape()
+
+        x = np.random.uniform(0.1, 1, self.shape).astype(self.dtype)
+        out = np.apply_along_axis(stable_softmax, 1,
+                                  x.reshape([-1, self.shape[-1]]))
+        out = out.reshape(self.shape)
 
-        x = np.random.uniform(0.1, 1, [10, 10]).astype(self.dtype)
-        out = np.apply_along_axis(stable_softmax, 1, x)
         self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
         self.outputs = {'Out': out}
         self.attrs = {
@@ -63,6 +70,11 @@ class TestSoftmaxOp(OpTest):
             self.check_grad(["X"], "Out", max_relative_error=0.01)
 
 
+class TestSoftmaxOp2(TestSoftmaxOp):
+    def get_x_shape(self):
+        return [2, 3, 4, 5]
+
+
 @unittest.skipIf(not core.is_compiled_with_cuda(),
                  "core is not compiled with CUDA")
 class TestSoftmaxCUDNNOp(TestSoftmaxOp):
@@ -70,6 +82,13 @@ class TestSoftmaxCUDNNOp(TestSoftmaxOp):
         self.use_cudnn = True
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
+class TestSoftmaxCUDNNOp2(TestSoftmaxCUDNNOp):
+    def get_x_shape(self):
+        return [2, 3, 4, 5]
+
+
 @unittest.skipIf(not core.is_compiled_with_cuda(),
                  "core is not compiled with CUDA")
 class TestSoftmaxFP16Op(TestSoftmaxOp):
@@ -83,6 +102,13 @@ class TestSoftmaxFP16Op(TestSoftmaxOp):
                 self.check_output_with_place(place, atol=1e-3)
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
+class TestSoftmaxFP16Op2(TestSoftmaxFP16Op):
+    def get_x_shape(self):
+        return [2, 3, 4, 5]
+
+
 @unittest.skipIf(not core.is_compiled_with_cuda(),
                  "core is not compiled with CUDA")
 class TestSoftmaxFP16CUDNNOp(TestSoftmaxOp):
@@ -97,10 +123,22 @@ class TestSoftmaxFP16CUDNNOp(TestSoftmaxOp):
                 self.check_output_with_place(place, atol=1e-3)
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
+class TestSoftmaxFP16CUDNNOp2(TestSoftmaxFP16CUDNNOp):
+    def get_x_shape(self):
+        return [2, 3, 4, 5]
+
+
 class TestSoftmaxMKLDNNOp(TestSoftmaxOp):
     def init_kernel_type(self):
         self.use_mkldnn = True
 
 
+class TestSoftmaxMKLDNNOp2(TestSoftmaxMKLDNNOp):
+    def get_x_shape(self):
+        return [2, 3, 4, 5]
+
+
 if __name__ == "__main__":
     unittest.main()