Hello i have a question of PeopleNet

Hi i using the jetson nano & DS 5.0

I don’t know much about tlt.

I’d like to retrain PeopleNet, remove the other classes, and use only the Person class.
Can I raise the frame rate higher than 10 when I retrain by using tlt?
I want to be guaranteed more than 20 frames at 360p.
If that is possible, I would like to try the tlt.

If you have this kind of tutorial, please share it.

thankyou !..

It should be possible.
My reason:

  1. you only train one class.
  2. you can prune more to get a smaller trt engine.
    Reference: Accelerating Peoplnet with tlt for jetson nano - #13 by Morganh
    In that link, see “Pruning ratio (pruned model / original model): 0.143”. The pruned tlt model is 14.3% of unpruned tlt model. You can prune more.

More, the peoplenet is trained with resnet34 backbone. If change to resnet18, the fps will be larger.

Hi @Morganh

thank your comment!

I have been using resnet10 until now, but often the model recognized the black chair as a person
,even PeopleNet sometimes did. Can i reduce this if i retraining it?

And I want to buy a learning PC. I’m not sure what level of specification I need. Is there any problem with learning even if there is a single GPU? (Like single RTX2080)

could you recommend??

1)Have you finished training with resnet10 backbone? What is the mAP result? If possible, please paste your training spec.

2)For PC, please see Integrating TAO Models into DeepStream — TAO Toolkit 3.22.05 documentation

3)It is fine to train with one GPU.

  1. That mean, i have been using built-in model resnet10(Primary_Nano) & pretrained_tlt_peoplenet in deepstream5.0

  2. There is no name for the minimum recommended product.

  • 4 GB system RAM
  • 4 GB of GPU RAM
  • Single core CPU
  • 1 GPU
  • 50 GB of HDD space
  1. Are you saying that just one GPU will enough speed and performance to retrain PeopleNet?
  1. Could you tell me what is the built-in model resnet10(Primary_Nano)?
  2. Please note that you should use your host PC to run tlt training. That is Hardware Requirements of your host PC. Only the GPU is the NV’s product.
  3. One GPU is enough for tlt training. If you have multi gpus, the training speed will speed up.

Question 1 is not very important.
(that’s mean I used the basic model provided DS5.0 DP version
path(resnet) = \opt\nvidia\deepstream\deepstream-5.0\samples\models\Primary_Detector_Nano
path(resnet) = \opt\nvidia\deepstream\deepstream-5.0\samples\models\tlt_pretrained_models)

below is peoplenet on DS 5.0 DP ver

I want to reduce the probability of recognizing chairs as people, as in the picture above.
and like my first question.
If I only want to use Person Class, can I reduce Layer to 18 levels and compress it to achieve more than 20 fps performance?

and

And it is the specification of host PC that I am thinking of, please confirm it.

cpu - amd 3900x
ram - 32gb
ssd - 500g
hdd - 2tb
gpu - (could you recommend a suitable specification for my retrain???) ex rtx2080 x2 or titan ~~blahblah

  1. I think your path is not correct for running peoplenet.
    Please use below link.
    /opt/nvidia/deepstream/deepstream-5.0/samples/configs/tlt_pretrained_models
    Firstly, please read the README, then try to run the peoplenet. All the config file is ready by default.

For fps, please refer to my previous comment. It should be possible.

  1. The config of your host PC is fine. More, titan or trx2080 is also ok.

Thank you @Morganh

i have one more question!

  1. When attempting to configure GPU on a host PC with two RTX2070 Super, Is nvlink required when using tlt? Or does it just increase performance by itself?

  2. Can you recommend it?

  • one RTX2080 ti
  • two RTX2070 Super
  • two RTX2070 Super + nvlink (in Q1 - If ‘tlt’ requires ‘nvlink’ when using two GPUs,)
  1. Running tlt does not depend on nvlink.
  2. all should be working.

Hi @Morganh

Thank you so much for answering so quickly every time.

I will retrain as you advised and ask you again.