bug fixes for new MirrorPadGrad NMS IOU comparison

mindspore/ccsrc/backend/kernel_compiler/gpu/cuda_impl/mirror_pad_impl.cu

@@ -100,7 +100,7 @@ __global__ void MirrorPadGrad(const size_t size, const T *dy, const int num, con
     const int padded_y = ((pos / old_width) % old_height) + pad_top_;
 
     // copy positions own value into output
-    dx[pos] = dx[pos] + dy[(block_num * padded_height + padded_y) * padded_width + padded_x];
+    dx[pos] = dy[(block_num * padded_height + padded_y) * padded_width + padded_x];
 
     int x_dist_1 = (ap1_x - padded_x - adjust);
     int y_dist_1 = (ap1_y - padded_y - adjust);

mindspore/ccsrc/backend/kernel_compiler/gpu/cuda_impl/nms_with_mask_impl.cu

@@ -89,7 +89,7 @@ __inline__ __device__ bool IOUDecision(T *output, int box_A_ix, int box_B_ix, in
   T height = max(y_2 - y_1, T(0));
   T combined_area = area[box_A_ix] + area[box_B_ix];
   // return decision to keep or remove box
-  return !(((width * height) / (combined_area - (width * height))) >= IOU_value);
+  return !(((width * height) / (combined_area - (width * height))) > IOU_value);
 }
 
 // calculate areas for boxes -> sorted by output boxes