TDA4VM: Query regarding CenterNet model porting

Gina Mathew

Prodigy 140 points

Part Number: TDA4VM

Hi,

As part of Customer project we have to convert centernet pytorch model to TIDL format.

Steps followed

1.Converted Pytorch model to ONNX

2. Executed TIDL Import tool and got TIDL binary files.

TIL Binaries were generated, but the text file generated (expected to contain the bounding box coordinate) corresponding to the given input image is empty. Also the post processed image generated doesn't contain bounding box overlaid.

We have given postProcType =2 in configuration file. Since no bounding box are not generated we have couple of queries in that. We analyzed the network architecture generated. Currently onnx model give output from the 3 centernet heads- Heatmap head, Dimension head and Offset head. While converting this model to TIDL format, we could find the meta-architecture support for Object Detection for these architectures - SSD, Yolo and Retinanet. But we would like to know whether meta architecture support is provided for centernet? Can you please provide inputs regarding this

regards,

Gina

over 4 years ago

0 kumar.desappan over 4 years ago

TI__Mastermind 22145 points

As of now, we support only SSD, Yolo and Retinanet post-processing in C7X-DSP. Any other OD post-processing can be support on ARM A72. please use ONNX runtime APIs to run these models on TDA4VM and run the post-processing using standard c++/python code similar to the one used in your training code. Measure the time/latency of this post-processing in ARM, if this is not meeting the application requirement, please do let us know the latency details.

0 kumar.desappan over 4 years ago in reply to kumar.desappan

TI__Mastermind 22145 points

Refer below for ONNXRT API based Python and CPP examples

github.com/.../edgeai-tidl-tools

0 Gina Mathew over 4 years ago in reply to kumar.desappan

Prodigy 140 points

Hi,

Thanks for your reply. We are using RTOS SDK version 07.01.00.11. My understanding is that 7.1 version doesn't support ONNX runtime support with heterogenous support. So we will have to use TIDL-RT and implement the post-processing part in C and run on ARM A72. Is my understanding correct.

regards,

Gina

0 kumar.desappan over 4 years ago in reply to Gina Mathew

TI__Mastermind 22145 points

Yes, your understanding is right.

0 Gina Mathew over 4 years ago in reply to kumar.desappan

Prodigy 140 points

Hi,

We will implement the post processing and let you know regarding the performance.

regards,

Gina

Processors

Processors forum

TDA4VM: Query regarding CenterNet model porting