Onnx to tensorrt engine optimization profile error

[ozgyal]

Hello, thanks for sharing your code. I'm getting this error while converting the created onnx (with export_onnx.py) to engine even though orig_im_size is not in dynamic_axes: Internal Error (orig_im_size: optimization profile is missing values for shape input)

Any thoughts that might help on this issue?

Thank you.

[ryouchinsa]

Thanks for writing the issue.

Do you mean that this command worked in your environment and you obtained the ONNX model?
python export_onnx.py --mode export

Then, in which command was the error message shown?

[ozgyal]

Thanks for getting back and sorry for not being clear. Actually, the error happens while converting onnx to tensorrt engine in a c++ environment. It could be reproduced with trtexec --onnx=<name_of_onnx>.onnx command as well. It may be an issue due to the tensorrt version but I'm not sure. There is an ambiguity for the input orig_im_size. Tensorrt says it should have a dynamic shape.

But I guess I have solved it by removing orig_im_size from the model inputs since it is a constant value as (1024, 1024) (right?). I've just put the dimensions into the model. Now I can get the engine.

[CvBokchoy]

谢谢你回来，抱歉我说得不清楚。实际上，在c++环境中将onnx转换为tensorrt引擎时会发生错误。也可以使用trtexec --onnx=<name_of_onnx>.onnx命令复制它。这可能是由于tensorrt版本的问题，但我不确定。输入orig_im_size存在歧义。Tensorrt说它应该有一个动态的形状。

但是我想我已经通过从模型输入中删除orig_im_size解决了这个问题，因为它是一个常量值，为（1024，1024）（对吗？）。我刚把尺寸输入模型。现在我可以得到引擎。

Hello, I converted two onnx models using the code provided by the author. Encoder model and decoder model, in which the encoder model can be successfully converted to tensorRT engine, but the decoder has encountered problems, may I ask your decoder model in the conversion of Engine encountered problems?

[ozgyal]

I used this monkey patch: xuanlinli17/efficientvit@322ce5e @CvBokchoy

[CvBokchoy]

我用了这个猴子补丁：xuanlinli 17/efficientvit@ 322ce5e @CvBokchoy

Hello, I am not good at checking in this way, can you send me the correct pt->onnx code that can convert trt by email? My email address is 3388530136@qq.com （你好，我不太会用这种方式查看，您能将正确可以转换trt的 pt->onnx的代码 以邮件的方式发送给我吗? 我的邮箱是3388530136@qq.com）

[CvBokchoy]

I used this monkey patch: xuanlinli17/efficientvit@322ce5e @CvBokchoy

Please send me the complete pt->onnx and the conversion code that can convert encode decode model to tensorRT engine by email. Thank you（请您将完整的pt->onnx 并且可以将encode decode模型转换为tensorRT engine 的转换代码，以邮件的方式发送给我 谢谢您）