diff --git a/paddle/fluid/operators/label_smooth_op.cc b/paddle/fluid/operators/label_smooth_op.cc
index b76ccbcfce09e1aea933f0a909530f4fd5e22a3e..f47233b6e1b38f0f90210e7d9532ff1c816574fd 100644
--- a/paddle/fluid/operators/label_smooth_op.cc
+++ b/paddle/fluid/operators/label_smooth_op.cc
@@ -37,7 +37,7 @@ class LabelSmoothOp : public framework::OperatorWithKernel {
       auto noise_dims = ctx->GetInputDim("PriorDist");
       auto noise_numel = paddle::framework::product(noise_dims);
       PADDLE_ENFORCE(
-          in_dims[1] == noise_numel,
+          in_dims[in_dims.size() - 1] == noise_numel,
           "The number of elements in Input(PriorDist) must be equal to the "
           "dimension of each label.");
     }
diff --git a/paddle/fluid/operators/label_smooth_op.cu b/paddle/fluid/operators/label_smooth_op.cu
index 89f1d28e9988281c77e0cefa349bd7181b432c20..33ae35a81f848129223815c5c08f5de82b329f92 100644
--- a/paddle/fluid/operators/label_smooth_op.cu
+++ b/paddle/fluid/operators/label_smooth_op.cu
@@ -34,7 +34,7 @@ __global__ void LabelSmoothRunDistKernel(const int N, const float epsilon,
                                          const T* dist_data, T* dst) {
   int idx = blockDim.x * blockIdx.x + threadIdx.x;
   for (; idx < N; idx += blockDim.x * gridDim.x) {
-    int dist_idx = idx - (idx / dist_numel) * dist_numel;
+    int dist_idx = idx % dist_numel;
     dst[idx] = static_cast<T>(1 - epsilon) * src[idx] +
                static_cast<T>(epsilon) * dist_data[dist_idx];
   }
@@ -56,7 +56,7 @@ class LabelSmoothGPUKernel : public framework::OpKernel<T> {
     auto* out_t = ctx.Output<framework::LoDTensor>("Out");
     auto* in_t = ctx.Input<framework::LoDTensor>("X");
     auto* dist_t = ctx.Input<framework::Tensor>("PriorDist");
-    auto label_dim = in_t->dims()[1];
+    auto label_dim = in_t->dims()[in_t->dims().size() - 1];
     auto epsilon = ctx.Attr<float>("epsilon");
     auto& dev = *ctx.template device_context<DeviceContext>().eigen_device();
     auto size_prob = in_t->numel();
diff --git a/paddle/fluid/operators/label_smooth_op.h b/paddle/fluid/operators/label_smooth_op.h
index f3da17de011053fa118b5a4257bb5c3b00084741..760d542505ec138eca032b82a7e7902db9c0887a 100644
--- a/paddle/fluid/operators/label_smooth_op.h
+++ b/paddle/fluid/operators/label_smooth_op.h
@@ -27,7 +27,7 @@ class LabelSmoothKernel : public framework::OpKernel<T> {
     auto* out_t = ctx.Output<framework::LoDTensor>("Out");
     auto* in_t = ctx.Input<framework::LoDTensor>("X");
     auto* dist_t = ctx.Input<framework::Tensor>("PriorDist");
-    auto label_dim = in_t->dims()[1];
+    auto label_dim = in_t->dims()[in_t->dims().size() - 1];
     out_t->mutable_data<T>(ctx.GetPlace());
 
     auto epsilon = ctx.Attr<float>("epsilon");
@@ -39,7 +39,7 @@ class LabelSmoothKernel : public framework::OpKernel<T> {
       out.device(dev) =
           static_cast<T>(1 - epsilon) * in +
           static_cast<T>(epsilon) *
-              dist.broadcast(Eigen::DSizes<int, 1>(in_t->numel()));
+              dist.broadcast(Eigen::DSizes<int, 1>(in_t->numel() / label_dim));
     } else {
       out.device(dev) = static_cast<T>(1 - epsilon) * in +
                         static_cast<T>(epsilon / label_dim);
diff --git a/python/paddle/fluid/tests/unittests/test_label_smooth_op.py b/python/paddle/fluid/tests/unittests/test_label_smooth_op.py
index 0103b86491266d9fb7c2a09d7b7b6fe89dcba055..222e1321fecf8a383ebdba66ff17e760bab6ae51 100644
--- a/python/paddle/fluid/tests/unittests/test_label_smooth_op.py
+++ b/python/paddle/fluid/tests/unittests/test_label_smooth_op.py
@@ -53,5 +53,23 @@ class TestLabelSmoothOpWithPriorDist(TestLabelSmoothOp):
         self.outputs = {'Out': smoothed_label}
 
 
+class TestLabelSmoothOp3D(TestLabelSmoothOp):
+    def setUp(self):
+        super(TestLabelSmoothOp3D, self).setUp()
+        self.inputs['X'] = self.inputs['X'].reshape(
+            [2, -1, self.inputs['X'].shape[-1]])
+        self.outputs['Out'] = self.outputs['Out'].reshape(self.inputs['X']
+                                                          .shape)
+
+
+class TestLabelSmoothOpWithPriorDist3D(TestLabelSmoothOpWithPriorDist):
+    def setUp(self):
+        super(TestLabelSmoothOpWithPriorDist3D, self).setUp()
+        self.inputs['X'] = self.inputs['X'].reshape(
+            [2, -1, self.inputs['X'].shape[-1]])
+        self.outputs['Out'] = self.outputs['Out'].reshape(self.inputs['X']
+                                                          .shape)
+
+
 if __name__ == '__main__':
     unittest.main()