专栏算法工具链LaneATT模型推理出错The model [xxx] genarate 1602 funccalls, but system only support 1024 funccalls at most

LaneATT模型推理出错The model [xxx] genarate 1602 funccalls, but system only support 1024 funccalls at most

已解决
mq2023-05-17
19
4
  1. 系统软件版本:x3_ubuntu_v1.1.3

  2. 问题描述:

模型是LaneATT(https://github.com/lucastabelini/LaneATT),转换后在x3pi上执行,出现如下错误:

[E][DNN][multi_model_task.cpp:247](1684315378625) The model [LaneATT_test_rgb_2core] genarate 1602 funccalls, but system only support 1024 funccalls at most. Funccalls can be reduced by: 1. Increase `max-time-per-fc` time at compile stage; 2. Decrease the number of batchsize at compile stage; 3. Reduce the number of roi when calling hbDNNRoiInfer

[W][DNN]executor.cpp:136](1684315378625) ri_id=1 is running when invoke hrRiDestroy!

[E][DNN][multi_model_task.cpp:1216](1684315378625) RiDestroy failed!

错误结果中给出的修改建议:

1. Increase `max-time-per-fc` time at compile stage; 试了没用

2. Decrease the number of batchsize at compile stage; batchsize是1

3. Reduce the number of roi when calling hbDNNRoiInfer; 不知道怎么操作

模型(链接:https://pan.baidu.com/s/1F3gYd8TMZtblVYQuj0632Q 提取码:xsig )

算法工具链
评论2
0/1000
  • 颜值即正义
    Lv.2

    这里注意到您模型后面有非常多的CPU算子,这种情况下不适合做板端部署。另外通过新版本OE包提供的hrt_model_exec检验发现,是子图1产生的funccalls过多,超出了系统软件的上限。建议先尝试将CPU算子配置run_on_bpu,看看是否正常,如果依然有问题建议对模型进行修改。

    2023-05-18
    0
    2
  • 颜值即正义
    Lv.2
    2023-05-18
    0
    0