ONNX Output format

[andrzejom]

❔Question

I have a question regarding the output that the converted model generates. When I inspect the model in netron looks like this:
https://imgur.com/yzGe39b
The last layer outputs the values into a Tensor of a size 1x25200x8. What does this vector contain? I'm looking into getting the values of Bounding boxes for detected classes however it seems a bit confusing. The model has 3 classes and the output from the last layer looks like this
https://imgur.com/LDTCPfR
It doesn't make a lot of sense to me. Perhaps someone could enlighten me

Additional context

ONNX 1.9.0
Torch 1.8.1

[leoll2]

It's the output of the Detect layer.

1 -> batch size
25200 -> number of predictions
8 -> x y w h objectness conf_cls1 conf_cls2 conf_cls3

You probably want to apply NMS to the predictions, which you have to apply as a post-processing stage because it's not part of the exported onnx model.

[iperov]

export to onnx does not contain Detect layer
and the model output 1,3,80,80,16

so how to interpret these values without Detect layer on pure python?

[andrzejom]

If you choose to export model without detect layer it is indeed true that you will end up with the raw output ' 1,3,80,80,16 '. In this case you need to write your extra detect layer on top over it

[iperov]

you need to write your extra detect layer on top over it

are you serious ?

[iperov]

who want to dig into that dirty code and write DetectLayer from scratch?

why you don't provide pure python/numpy code to process raw output from onnx model ?

[andrzejom]

But why you want to process raw output in the first place? You can use predefined detector with the architecture

[iperov]

because I want to use onnx model where pytorch will not be used

[iperov]

did you hear about simplicity and minimalism?
It is not about Yolo repositories :D

[iperov]

nevermind, I investigated the code and writing the layer

[glenn-jocher]

ONNX exports include Detect() layer.

[PaginDm]

nevermind, I investigated the code and writing the layer

So I have the same problem. I'm trying to write custom postprocessing for native 1,3,80,80,16 layer, but WTF. This output is different from YOLOv3 in Tensorflow implementations for example. Therefore my boxes are wrong. Can you share Detect-layer implementation?

[iperov]

only the code of Detect Layer will not help.
I changed internal code of the repo to export right onnx model which works without warnings and error at max speed.

it will be shared with DeepFaceLive release.

[AdrianOverture]

ONNX exports include Detect() layer.

I exported a yolov5s trained with a custom dataset using the export.py script. But I cannot manage to export it with the Detect layer. How can I do it? @glenn-jocher

[glenn-jocher]

@AdrianOverture exported models automatically include detect layer.