--> Op 1                
      -----/      |                 
   --/            v                 
 NCHW           Out 1 --------> NHWC
   --\            |                 
      -----\      v                 
            --> Op 2     

(%1069) = "sigmoid(phi_kernel)" (%1068) {kernel_key:<backend:CPU|layout:NCHW|dtype:float32>,kernel_name:"sigmoid",op_name:"pd_op.sigmoid",origin_id:1936,stop_gradient:[false],struct_name:"/CenterNetHead/"} : (cpu_tensor<-1x80x128x128xf32>) -> cpu_tensor<-1x80x128x128xf32>
(%1070) = "transpose(phi_kernel)" (%1069) {kernel_key:<backend:CPU|layout:NCHW|dtype:float32>,kernel_name:"transpose",op_name:"pd_op.transpose",origin_id:1937,perm:[0,2,3,1],source:"transfer_layout_pass",stop_gradient:[false]} : (cpu_tensor<-1x80x128x128xf32>) -> cpu_tensor<-1x128x128x80xf32>
......
(%1082) = "transpose(phi_kernel)" (%1081) {kernel_key:<backend:CPU|layout:NHWC|dtype:float32>,kernel_name:"transpose",op_name:"pd_op.transpose",origin_id:1949,perm:[0,3,1,2],source:"transfer_layout_pass",stop_gradient:[false]} : (cpu_tensor<-1x128x128x80xf32>) -> cpu_tensor<-1x80x128x128xf32>
(%1083) = "equal(phi_kernel)" (%1082, %1070) {kernel_key:<backend:CPU|layout:NCHW|dtype:float32>,kernel_name:"equal",op_name:"pd_op.equal",origin_id:1950,stop_gradient:[true]} : (cpu_tensor<-1x80x128x128xf32>, cpu_tensor<-1x128x128x80xf32>) -> cpu_tensor<-1x80x128x128xb>

(%1070) = "sigmoid(phi_kernel)" (%1069) {kernel_key:<backend:npu|layout:NCHW|dtype:float32>,kernel_name:"sigmoid",op_name:"pd_op.sigmoid",origin_id:1935,stop_gradient:[false],struct_name:"/CenterNetHead/"} : (custom_device_tensor<-1x80x128x128xf32>) -> custom_device_tensor<-1x80x128x128xf32>
(%1071) = "transpose(phi_kernel)" (%1070) {kernel_key:<backend:npu|layout:NCHW|dtype:float32>,kernel_name:"transpose",op_name:"pd_op.transpose",origin_id:1936,perm:[0,2,3,1],source:"transfer_layout_pass",stop_gradient:[false]} : (custom_device_tensor<-1x80x128x128xf32>) -> custom_device_tensor<-1x128x128x80xf32>
(%1072) = "transpose(phi_kernel)" (%1071) {kernel_key:<backend:npu|layout:NHWC|dtype:float32>,kernel_name:"transpose",op_name:"pd_op.transpose",origin_id:1937,perm:[0,3,1,2],source:"transfer_layout_pass",stop_gradient:[false]} : (custom_device_tensor<-1x128x128x80xf32>) -> custom_device_tensor<-1x80x128x128xf32>
......
(%1084) = "transpose(phi_kernel)" (%1083) {kernel_key:<backend:npu|layout:NHWC|dtype:float32>,kernel_name:"transpose",op_name:"pd_op.transpose",origin_id:1949,perm:[0,3,1,2],source:"transfer_layout_pass",stop_gradient:[false]} : (custom_device_tensor<-1x128x128x80xf32>) -> custom_device_tensor<-1x80x128x128xf32>
(%1085) = "equal(phi_kernel)" (%1084, %1072) {kernel_key:<backend:npu|layout:NCHW|dtype:float32>,kernel_name:"equal",op_name:"pd_op.equal",origin_id:1950,stop_gradient:[true]} : (custom_device_tensor<-1x80x128x128xf32>, custom_device_tensor<-1x80x128x128xf32>) -> custom_device_tensor<-1x80x128x128xb>