Bpnet input and output tensors

I am looking for documentation on the shapes and content of input and output tensors of bpnet for use for inference on tensorRT under C++

Thanks in advance for your help!

Actually you can use below way to check against any tensorrt engine.

$ python -m pip install colored
$ python -m pip install polygraphy --index-url https://pypi.ngc.nvidia.com
$ polygraphy inspect model xxx.engine

Thanks! Still trying to generate the engine… Once I have it I’ll try your suggestion

So I got the output tensor shapes as follows:

Binding Index: 0 (Input) [Name: input_1:0] | Shapes: min=(1, 288, 384, 3), opt=(4, 288, 384, 3), max=(16, 288, 384, 3)
Binding Index: 1 (Output) [Name: heatmap_out/BiasAdd:0] | Shape: (-1, 36, 48, 19)
Binding Index: 2 (Output) [Name: conv2d_transpose_1/BiasAdd:0] | Shape: (-1, 144, 192, 38)

I understand from this that there are two output tensors, a tensor of heatmaps which are actually Part Confidence Maps and the conv2d_transpose_1/BiasAdd corresponds to the Part Affinity Fields (PAFs).

Is there some documentation explaining the content of those two tensors?

You can refer to deepstream_tao_apps/bodypose2d_pgie_config.txt at master · NVIDIA-AI-IOT/deepstream_tao_apps · GitHub

Thanks for your reply which basically says:


Which does not provide any insight or help whatsoever in relation to my quesion.

There is no update from you for a period, assuming this is not an issue anymore.
Hence we are closing this topic. If need further support, please open a new one.

You can refer to BodyPoseNet | NVIDIA NGC
and Body Pose Estimation — TAO Toolkit 3.22.05 documentation

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.