diff --git a/paddle/fluid/operators/affine_grid_op.cu b/paddle/fluid/operators/affine_grid_op.cu
index 7aaaa0002c5ab31af72c75e69f5a283c09633ba4..58b56bdcf5614ed9183ce3bf11c1767f92650d20 100644
--- a/paddle/fluid/operators/affine_grid_op.cu
+++ b/paddle/fluid/operators/affine_grid_op.cu
@@ -62,11 +62,11 @@ __global__ void affine_grid_kernel(const int count, int n, int out_h, int out_w,
 
     int theta_offset = n * 6;  // 2 * 3;
     // affine from (h_coor, w_coor) to (x, y)
-    output[index * 2] = theta[theta_offset] * h_coor +
-                        theta[theta_offset + 1] * w_coor +
+    output[index * 2] = theta[theta_offset] * w_coor +
+                        theta[theta_offset + 1] * h_coor +
                         theta[theta_offset + 2];
-    output[index * 2 + 1] = theta[theta_offset + 3] * h_coor +
-                            theta[theta_offset + 4] * w_coor +
+    output[index * 2 + 1] = theta[theta_offset + 3] * w_coor +
+                            theta[theta_offset + 4] * h_coor +
                             theta[theta_offset + 5];
   }
 }
@@ -86,13 +86,13 @@ __global__ void affine_grid_grad_kernel(const int count, int n, int out_h,
 
     int theta_offset = n * 6;  // 2 * 3;
     T out_grad_x = out_grad[index * 2];
-    platform::CudaAtomicAdd(theta_grad + theta_offset, out_grad_x * h_coor);
-    platform::CudaAtomicAdd(theta_grad + theta_offset + 1, out_grad_x * w_coor);
+    platform::CudaAtomicAdd(theta_grad + theta_offset, out_grad_x * w_coor);
+    platform::CudaAtomicAdd(theta_grad + theta_offset + 1, out_grad_x * h_coor);
     platform::CudaAtomicAdd(theta_grad + theta_offset + 2, out_grad_x);
 
     T out_grad_y = out_grad[index * 2 + 1];
-    platform::CudaAtomicAdd(theta_grad + theta_offset + 3, out_grad_y * h_coor);
-    platform::CudaAtomicAdd(theta_grad + theta_offset + 4, out_grad_y * w_coor);
+    platform::CudaAtomicAdd(theta_grad + theta_offset + 3, out_grad_y * w_coor);
+    platform::CudaAtomicAdd(theta_grad + theta_offset + 4, out_grad_y * h_coor);
     platform::CudaAtomicAdd(theta_grad + theta_offset + 5, out_grad_y);
   }
 }
diff --git a/python/paddle/fluid/tests/unittests/test_affine_grid_op.py b/python/paddle/fluid/tests/unittests/test_affine_grid_op.py
index 55612d71a17a7ae9801535bf5a35c83b100aab30..d3e990ca13eb2911ea04ed546b91f58e2db4e440 100644
--- a/python/paddle/fluid/tests/unittests/test_affine_grid_op.py
+++ b/python/paddle/fluid/tests/unittests/test_affine_grid_op.py
@@ -49,7 +49,6 @@ class TestAffineGridOp(OpTest):
         self.initTestCase()
         self.op_type = "affine_grid"
         theta = np.random.randint(1, 3, self.theta_shape).astype("float32")
-        theta = np.ones(self.theta_shape).astype("float32")
         self.inputs = {'Theta': theta}
         self.attrs = {
             "use_cudnn": self.use_cudnn,