how to convet torch to onnx and then tensorrt

[kkeono2]

Thank you for your great work.

Do you have any plan to support model conveting implementation for onnx ?

[codet-ovd]

Hi,

You can use detectron2's built-in script to export the model to onnx. Also, since this codebase is built upon Detic, I believe related discussions in their Issue #61 and #107 may help you.

[gigasurgeon]

you can look at this discussion -> facebookresearch/Detic#113 (comment)

[kkeono2]

Thanky you for rapid answering @codet-ovd

I used detectron2's built-in script (i just change set-up configuration in export_onnx.py ) to convert torch model to onnx and i have successfully converted.
However, i faced the problem with topk initialization when i converted onnx to tensorrt as below(TensorRT error message).
'''
[6] Invalid Node - /proposal_generator/TopK
This version of TensorRT only supports input K as an initializer
'''

I guess the problem with topk initialization due to varing the image_thresh per image in centernet.py.
Therefore, i fixed the image_thresh as constant value(0.1).

            if num_dets > post_nms_topk:
                cls_scores = result.scores
                # image_thresh, _ = torch.kthvalue(cls_scores.float().cpu(),num_dets - post_nms_topk + 1)
                keep = cls_scores >= 0.1 #image_thresh.item()  

I also faced the error as below. i don't know why these error is occuring.
'''
[6] Invalid Node - /roi_heads/box_pooler/level_poolers.0/If
/roi_heads/box_pooler/level_poolers.0/If_OutputLayer: IIfConditionalOutputLayer inputs must have the same shape.
'''

Can you help me solve this please?
I use TensorRT 8.5.1, ONNX 1.12.0

[kkeono2]

Thank you for answering @gigasurgeon

I can convert torch model to onnx without comment the nms_and_topk line in centernet, while exporting the model.

Actually, i can't convet onnx to tensorrt as above issue.

Can you give me a suggetion to solve above issue ?