Inference on custom RTDETR network

[immohan6598]

Hi,

Thanks for sharing a work!

I have trained a custom model using https://github.com/lyuwenyu/RT-DETR/tree/main/rtdetr_pytorch

network dimension is 1024 x 1024 and converted the .pt to tensorrt engine.

when i inference i get segmentation fault in the postprocessing

i only have 15 classes

please help me figure out this

[olibartfast]

Hi there, In my project I used a pre-trained RT-DETR from ultralytics so the network output layer might be different from your model. Also I used 640x640 hardcoded input, I should modify my project to make it work for any size (maybe specifying it from cli).

Which RT-DETR did you use from lyuwenyu model zoo?

What did you use to convert the binary to tensorrt engine, trtexec after onnx conversion or torch2trt?

[mohanrobro]

Hi, Thanks for the answer. I used this official repo https://github.com/lyuwenyu/RT-DETR/tree/main/rtdetr_pytorch Guideline is also available to convert a model to tensor-rt

I changed the hard-coded network dimension values.

I used this mode rtdetr_r18vd

I visualized the onnx model the output. I don't have that deep understanding on infer() function that updates ouputs and shapes so just want to understand.

I assume the infer will give the final output layer data.

Thanks in Advance.

[olibartfast]

Thanks all right, I'll take a look at it as soon as I can 👍

[immohan6598]

output layer dimension

[olibartfast]

Yes is a completely different input/output format to the model I used for this project.

[immohan6598]

will you be able to support? do you work on freelancing projects? I am working with freelancers to finish projects

[olibartfast]

Yes but not too soon, I have a full-time job so I can work on my GitHub projects only in my spare time.

[olibartfast]

Sorry a bit late :smile: , I currently wrote the code to infer this model here