TDA4VM: Inference for Object Detection model trained for TDA4VM using Edgeai-modelmaker

Chaitanya Suryawanshi

Part Number: TDA4VM
Other Parts Discussed in Thread: DRA829, AM67A

Tool/software:

I have trained a yolox_s_lite (yolox_s_lite_640x640_20220221_checkpoint.pth) on custom dataset using Edgeai-modelmaker for 100 epochs, my desired EVM is TDA4VM

the model should have been compiled their itself with Edgeai-modelmaker but the error encountered after training was,

FileNotFoundError: [Errno 2] No such file or directory: '/home/chai/edgeai-tensorlab/edgeai-modelmaker/data/projects/classwise/run/20250821-175504/yolox_s_lite/compilation/pkg/artifacts.yaml'

So After training I got the .onnx and .prototxt file , then I compiled the model using edgeai-tidl-tools , now I want the inference on the video with this model to validate its detections, but I am not able find a proper way to have inference from a custom trained model.

I am trying on the edgeai-studio-model analyzer for the inference but I am not able to find a option to upload a custom trained model. So till now we haven't been able to test our custom trained yolox_s_lite model on video or image. can you help with some of the methods for inference in this case ?

1 month ago

0 Wen Li 1 month ago

TI__Expert 7331 points

Hello;

Thanks for the question.

You can build a ROTS SDK, and download your yolo model (the artifact of your model) to an EVM and verify it. The real time video playback on EVM is a good way to verify it.

There is an example in the SDK, you just need to modify the yaml file to invoke your model.

Make sure to download the test data tarball as well, there are video test streams there.

Here is the link to download SDK and how to run the TIDL vision applications.

https://www.ti.com/tool/PROCESSOR-SDK-J721E?keyMatch=TDA4VM%20sdk&tisearch=universal_search

Click this link, then, click the last download option "PROCESSOR-SDK-RTOS-J721E — RTOS SDK for DRA829 & TDA4VM Jacinto"

Best regards

Wen Li

0 Chaitanya Suryawanshi 1 month ago in reply to Wen Li

Prodigy 40 points

Hello,

Thank for the reply.

Is verifying the model on EVM is the only way for inference in this case?

Or else are there any ways through which we can have its inference on workstation/PC itself ?

Thanks

Chaitanya

0 Chaitanya Suryawanshi 1 month ago in reply to Wen Li

Prodigy 40 points

Hello,

EVM that you are specifying above is virtual EVM that you provide through model analyzer or it's actual TDA4VM board???

Thanks

Chaitanya

0 Chaitanya Suryawanshi 1 month ago in reply to Chaitanya Suryawanshi

Prodigy 40 points

hello,

any update on this ??

thanks.

0 Wen Li 1 month ago in reply to Chaitanya Suryawanshi

TI__Expert 7331 points

We recommend a real EVM, if you want to do a real time inferencing and verify the performance.

+1 Wen Li 1 month ago in reply to Wen Li

TI__Expert 7331 points

You can also use the "PC_dsp_test_dl_algo.out" program to do the inference test on PC. It is in the following directory:

../edgeai-tidl-tools/tools/AM67A/tidl_tools/

depend on the device you are using, your one may not be "AM67A".

You just need to create a inference config text file as an input to this program "PC_dsp_test_dl_algo.out"

You refer to this link for more details

https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/07_00_00_11/exports/docs/tidl_j7_01_02_00_09/ti_dl/docs/user_guide_html/md_tidl_sample_test.html

But, to use video streams as inputs to test inference on PC simulation will be slow and may involve some extra pre-processing (such as video decompress).

Again, we recommend using a real EVM, if you want to use video streams as inputs to test inference. Since the SDK has the infrastructure built in; video could be ether from camera or pre-recorded streams.

Best regards

Wen Li

0 Wen Li 1 month ago in reply to Wen Li

TI__Expert 7331 points

Hello;

Do you still have questions about this issue? If not, please close the ticket. You can re-open or submit a new one easily, if you have questions in the future.

Thanks and regards

Wen Li

0 Wen Li 1 month ago in reply to Wen Li

TI__Expert 7331 points

Hello;

If you really want to do simulation on PC, and if you have an image sensor (camera) which can capture a raw video clip. Then you can treat the clip as multiple frames/images (for your PC simulation), for example, 60 images for 2s video (assuming 30 frames/s). The minimum pre-processing you may to do is to convert the sensor format to RGB format, which is pretty simple.

Best regards

Wen Li

Processors

Processors forum

TDA4VM: Inference for Object Detection model trained for TDA4VM using Edgeai-modelmaker