生成的onnx模型在ai_toolchain中进行模型check检查时出现错误Segmentation fault

已解决

lujiaxin2022-12-22

生成好的onnx模型在使用 hb_mapper checker检查时出现Segmentation fault，开发者工具版本为horizon_xj3_open_explorer_v2.2.3_20220617，希望能帮忙解决

[root@293c13622561 codes]# hb_mapper checker --model-type onnx --march bernoulli2 --model best1.onnx

2022-12-21 19:23:37,154 INFO Start hb_mapper....

2022-12-21 19:23:37,154 INFO log will be stored in /data/horizon_x3/codes/hb_mapper_checker.log

2022-12-21 19:23:37,155 INFO hbdk version 3.27.1

2022-12-21 19:23:37,155 INFO horizon_nn version 0.12.8

2022-12-21 19:23:37,156 INFO hb_mapper version 1.5.7

2022-12-21 19:23:37,156 WARNING parameter [output] is deprecated

2022-12-21 19:23:37,327 INFO Model type: onnx

2022-12-21 19:23:37,328 INFO input names []

2022-12-21 19:23:37,329 INFO input shapes {}

2022-12-21 19:23:37,329 INFO Begin model checking....

2022-12-21 19:23:37,332 INFO [Wed Dec 21 19:23:37 2022] Start to Horizon NN Model Convert.

2022-12-21 19:23:37,333 INFO The input parameter is not specified, convert with default parameters.

2022-12-21 19:23:37,333 INFO The hbdk parameter is not specified, and the submodel will be compiled with the default parameter.

2022-12-21 19:23:37,334 INFO HorizonNN version: 0.12.8

2022-12-21 19:23:37,334 INFO HBDK version: 3.27.1

2022-12-21 19:23:37,335 INFO [Wed Dec 21 19:23:37 2022] Start to parse the onnx model.

2022-12-21 19:23:37,984 INFO ONNX model info:

ONNX IR version: 6

Opset version: 11

Input name: input, [1, 32, 256, 7]

2022-12-21 19:23:39,214 INFO [Wed Dec 21 19:23:39 2022] End to parse the onnx model.

2022-12-21 19:23:39,241 INFO Model input names: ['input']

2022-12-21 19:23:39,426 INFO Saving the original float model: ./.hb_check/original_float_model.onnx.

2022-12-21 19:23:39,442 INFO [Wed Dec 21 19:23:39 2022] Start to optimize the model.

Segmentation fault

附件:

v1/static/fileData/best1_20221222110310.zip

算法工具链

0/600

颜值即正义
Lv.2
您好，这是由于模型中包含if算子导致，目前if算子尚不支持哈
2022-12-22
0
5
- lujiaxin回复颜值即正义:
  您好，已经将全部if算子去除，还是报一样的错，是有其他算子不支持吗？
  [root@293c13622561 codes]# hb_mapper checker --model-type onnx --march bernoulli2 --model best2.onnx
  2022-12-23 11:07:27,844 INFO Start hb_mapper....
  2022-12-23 11:07:27,844 INFO log will be stored in /data/horizon_x3/codes/hb_mapper_checker.log
  2022-12-23 11:07:27,845 INFO hbdk version 3.27.1
  2022-12-23 11:07:27,845 INFO horizon_nn version 0.12.8
  2022-12-23 11:07:27,846 INFO hb_mapper version 1.5.7
  2022-12-23 11:07:27,846 WARNING parameter [output] is deprecated
  2022-12-23 11:07:28,053 INFO Model type: onnx
  2022-12-23 11:07:28,054 INFO input names []
  2022-12-23 11:07:28,056 INFO input shapes {}
  2022-12-23 11:07:28,056 INFO Begin model checking....
  2022-12-23 11:07:28,058 INFO [Fri Dec 23 11:07:28 2022] Start to Horizon NN Model Convert.
  2022-12-23 11:07:28,058 INFO The input parameter is not specified, convert with default parameters.
  2022-12-23 11:07:28,058 INFO The hbdk parameter is not specified, and the submodel will be compiled with the default parameter.
  2022-12-23 11:07:28,059 INFO HorizonNN version: 0.12.8
  2022-12-23 11:07:28,059 INFO HBDK version: 3.27.1
  2022-12-23 11:07:28,059 INFO [Fri Dec 23 11:07:28 2022] Start to parse the onnx model.
  2022-12-23 11:07:28,773 INFO ONNX model info:
  ONNX IR version: 6
  Opset version: 11
  Input name: input, [1, 32, 256, 7]
  2022-12-23 11:07:30,063 INFO [Fri Dec 23 11:07:30 2022] End to parse the onnx model.
  2022-12-23 11:07:30,094 INFO Model input names: ['input']
  2022-12-23 11:07:30,284 INFO Saving the original float model: ./.hb_check/original_float_model.onnx.
  2022-12-23 11:07:30,300 INFO [Fri Dec 23 11:07:30 2022] Start to optimize the model.
  Segmentation fault
  回复中我无法添加附件，如果方便的话可以加个微信把模型发您，帮我看一下嘛
  2022-12-23
  0
  回复
- 颜值即正义回复lujiaxin:
  可以添加右下角地平线小助手的微信，或者提供百度网盘链接
  2022-12-23
  2
  回复
- 颜值即正义回复lujiaxin:
  您好，发现您并没有把模型中所有if算子删除。
  
  此外，看到您的工具版本比较旧了，为了方便您的验证，打包了三个whl包，您可以通过链接下载：
  https://pan.horizon.ai/index.php/s/KSByraPEStAQka7
  在docker中通过pip install *.whl即可安装。
  2022-12-23
  0
  回复
- lujiaxin回复颜值即正义:
  您好，您说的if算子是指条件语句if吗
  2022-12-23
  0
  回复
- 颜值即正义回复lujiaxin:
  是指转换生成onnx中存在的if分支，可以通过开源可视化工具netron进行查看
  2022-12-23
  0
  回复