2 Errors occurred while Using Hailo SDK

General
,
1
  1. Error Occurred in progress of parsing onnx file

hive1
hive11920×1080 278 KB

hive2
hive21920×1080 280 KB

  1. i can’t see the performance figures in profiler

hailo profiler --hef my_model.hef my_model_compiled_model.har

profilerissue
profilerissue894×574 17.2 KB

how can i solve these problems?

2

( my environment )
OS version = ubuntu 22.04
PCIE Driver = 4.17v
Hailo sw suite = 4.17v
CPU = i7 - 13700
kernel version = 6.5.0-35-generic
Hailo8

The process was performed well, but the numbers are not comming out.
is there something i need to set up?

3

hey @rhymus314

To get performance you have to run

hailortcli run {.hef} { --batch-size 8 , optional }

Regards

4

thanks, i can see the performance when using hailortcli run

but, is there any way i can see performance in Profiler .html ?
( the picture i attached shows the figures
Latency = N/A , OP/s = N/A, Input BW = N/A , Output BW = N/A )
why does it show N/A?

5

btw, how to solve the first question?
i had an error while parsing our onnx file

netron_hvn
netron_hvn484×526 7.04 KB

parse
parse1844×896 226 KB

hailo parser onnx --start-node-names images --end-node-names 2334 2313 2312 --hw-arch hailo8 my_model.onnx

my model made using pytorch, and is a model that converts the currently working model to onnx.
how could i know what the problem is?

[info] CPU : x86_64, i3-12100
[info] OS : Ubuntu 22.04
[info] Memory : Total: 15GB, Available: 12GB
[info] Kernel 6.5.0-35-generic
[info] Hailo DFC Version : 3.27.0
[info] HailoRT & Driver Version : 4.17.0

6

It shows N\A, since there are some depency on your system CPU, RAM etc.

7

Can you tell if this model passes ONNX check?

8

i checked this part has been solved, thanks

9

yes, i just checked it again, this model passes onnx checker

10

I’m a bit confused :thinking:, are you good now?

11

if i send you the onnx file to email, can you check it?

1 Like
12

yes, sure, please send it.

1 Like
13

As a first step, I always try to simlify the model, especially good for large models:
onnxsim det_fox.onnx det_fox-sim.onnx

After this, it’s easier to inspect the model structure, and I could see that in the begining there are a bunch of unsupported ops, probably there to address variable input frame sizes:
image
So, I’ve set the start-node-names to the first conv of the net. Then, I’ve used the last conv layers in the bulk of the model as end-nodes. This is the full command:

hailo parser onnx det_fox-sim.onnx \
--end-node-names \
    /rpn/head/bbox_pred_4/Conv /rpn/head/bbox_pred_3/Conv \ 
    /rpn/head/bbox_pred_2/Conv /rpn/head/bbox_pred_1/Conv \
    /rpn/head/bbox_pred/Conv /rpn/head/cls_logits_4/Conv \
    /rpn/head/cls_logits_3/Conv /rpn/head/cls_logits_2/Conv \
    /rpn/head/cls_logits_1/Conv /rpn/head/cls_logits/Conv \
    /backbone/fpn/layer_blocks.3/layer_blocks.3.0/Conv \
    /backbone/fpn/layer_blocks.0/layer_blocks.0.0/Conv \
    /backbone/fpn/layer_blocks.1/layer_blocks.1.0/Conv \
    /backbone/fpn/layer_blocks.2/layer_blocks.2.0/Conv \ 
--start-node-name /backbone/body/conv1/Conv
1 Like
14

oh, i totally understand it
now, it’s working

thank you for your support :smiley: