Inavalid Input PAF dimensions

Accelerated Computing Intelligent Video Analytics TAO Toolkit
1

Hi there i have been asked to raise the questions on this forum and pelase help me out in this,I have fine tuned Bodyposenet model on coco 2017 dataset and iam having issues that can be explroed in below conversation i had.-Link attached at the end.
Please let me know why and what happened wrong as i followed tao training notebook of bpnet i adjusted few of my parameters yes but the architechture of model is really changed?
Please review detailed conversation below from start in order to provide me a detailed resolution to this issue.

Link to Converstation

3

How did you finetune the bpnet model? Which pretrained model in ngc did you use?

4

HI there i used this model for fine-tuning

image
image796×150 14.2 KB

Also iam working with tao toolkit and notebook they have provided

5

Any updates please??

6

Hi,
I just search “invalid paf” and find a similar topic previously. See BodyPoseNet example on Jetson nano invalid input pafmap dimension - #2 by Morganh. Could you take a look?

Please

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/nvidia/deepstream/deepstream/lib/cvcore_libs
7

Nope this doesnt work for me still.

8

Please

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/nvidia/deepstream/deepstream/lib/cvcore_libs
9

i have done this please make this thing clear that i have fine tuned my model and then these issues are arising i have tested out default model it works fine

10

Thanks for the info. So, it works fine with below model , right?

image
image1577×127 54.2 KB

11

This is trainable model that has been used for finetuning,the default model i mean the onnx deployable model i tested out on pipeline.

12

OK. Can you test ngc deployable model as well? See BodyPoseNet | NVIDIA NGC.

13

Yes it works i have tested this

14

So, can you open the onnx files to compare and find the reason?

  • ngc onnx file (can work)
  • your onnx file after finetuning and exporting (cannot work)
15

i did the issue with fine tuning is of invalid input paf dimension

16

So, please change the dimension accordingly based on your training setting and exporting setting.

17

i have fine tuned it according to 17 coco keypoints and paf dimension has been set to 32 and for heatmap it is set to 18

18

Can you check the outputs?
For the ngc model, it is

image
image1151×703 32 KB

19

image
image509×527 27.8 KB

This is my fine tuned model output layers.
Also please see attached model file below:
bpnet_model.deploy.zip (60.2 MB)

20

So, please modify your output setting in the deepstream config file.
Please refer to deepstream_tao_apps/configs/nvinfer/bodypose2d_tao/bodypose2d_pgie_config.yml at release/tao5.3_ds7.0ga · NVIDIA-AI-IOT/deepstream_tao_apps · GitHub.

21

HI there i dont need to update the ouput blob names as they are same as default that has already been mentioned there.
Please provide accurate response.