what is the meaning of outputs?

[AliceSchaw]

I trained my custom dataset and exported the onnx model. the final names of output are 325 and 328, the type of them is float32[12348,4]. what is the meaning of outputs? my number of classes is 4.

onnx model:

cfg files:
prune_0.93_keep_0.01_16_shortcut_yolov3-spp3-4cls.txt

[github-actions]

Hello @AliceSchaw, thank you for your interest in our work! Please visit our Custom Training Tutorial to get started, and see our Google Colab Notebook, Docker Image, and GCP Quickstart Guide for example environments.

If this is a bug report, please provide screenshots and minimum viable code to reproduce your issue, otherwise we can not help you.

[glenn-jocher]

@AliceSchaw classes and boxes. boxes are nx4

[ghost]

@glenn-jocher Could you elaborate on how to interpret these values? I find them confusing as they are different from what the PyTorch version would output

[glenn-jocher]

@lostspirit0 these are designed for input into our iOS iDetection app.

You can export the raw model outputs as well, search the repo issues, there are many solutions already posted.

[AliceSchaw]

@lostspirit0 these are designed for input into our iOS iDetection app.

You can export the raw model outputs as well, search the repo issues, there are many solutions already posted.

@glenn-jocher Could you give me a link, thank you?

[glenn-jocher]

https://github.com/ultralytics/yolov3/issues?q=onnx+export

[AliceSchaw]

@glenn-jocher thanks for your reply. but i still do not understand the outputs, my input is 448X448, and strides are[32, 16, 8], so i guess the size of boundign box is (14 x 14 + 28 x 28 + 56 x 56) x 3 = 12348. but I want to get all the values that include the bbox of x, y, width, height and score, so, the expected size of outputs should be (14 x 14 + 28 x 28 + 56 x 56) x (4 + 1 + 4(class size)) * 3, but why is the size of outputs of the onnx 4 x 12348?

[glenn-jocher]

onnx model may have multiple outputs, you might want to visualize with netron.

[AliceSchaw]

onnx model may have multiple outputs, you might want to visualize with netron.

Yes, this is the visuality of my onnx model.

Can you give me some examples how to get all the values that include the bboxs of x, y, width, height and class score ?

[glenn-jocher]

@AliceSchaw ah perfect. Your outputs are right there. They're clearly labelled as 'boxes' and 'classes'. What seems to be the problem?

[AliceSchaw]

@AliceSchaw ah perfect. Your outputs are right there. They're clearly labelled as 'boxes' and 'classes'. What seems to be the problem?

here is official example: https://github.com/onnx/onnx-docker/blob/master/onnx-ecosystem/inference_demos/yoloV3_object_detection_onnxruntime_inference.ipynb

It is obviously different between the official example's onnx model and ultralytics's one.

The onnx model of official example has 3 outputs:
boxes: (1 x n_candidates x 4), the coordinates of all anchor boxes
scores: (1 x 80 x n_candidates), the scores of all anchor boxes per class
indices: (n_box x 3), selected indices from the boxes tensor. The selected index format is (batch_index, class_index, box_index)

However, here exported model has 2 outputs, any code example about utilizing the ultralytics's exported onnx model to detect objects with onnx runtime?

[glenn-jocher]

@AliceSchaw Ah I see. This repo's onnx integrates objectness and classification togethor, which is a common format used in mobile device detection, such as in our iOS iDetection app.

I don't think anyone actually uses onnx to run models, it's mostly just an interchange format to send your model elsewhere, i..e coreml, tflite, etc.

If you want you can try this:
https://pytorch.org/tutorials/advanced/super_resolution_with_onnxruntime.html#running-the-model-on-an-image-using-onnx-runtime

[AliceSchaw]

@AliceSchaw Ah I see. This repo's onnx integrates objectness and classification togethor, which is a common format used in mobile device detection, such as in our iOS iDetection app.

I don't think anyone actually uses onnx to run models, it's mostly just an interchange format to send your model elsewhere, i..e coreml, tflite, etc.

If you want you can try this:
https://pytorch.org/tutorials/advanced/super_resolution_with_onnxruntime.html#running-the-model-on-an-image-using-onnx-runtime

Thank you for your help and patience.

[leeyunhome]

@AliceSchaw classes and boxes. boxes are nx4

Hello.

Does the expression nx4 mean n * 4?

Thank you.

[glenn-jocher]

nx4 means shape = (n, 4)

[leeyunhome]

onnx model may have multiple outputs, you might want to visualize with netron.

Yes, this is the visuality of my onnx model.

Can you give me some examples how to get all the values that include the bboxs of x, y, width, height and class score ?

Hello,

#1109 (comment)

Thanks for leaving any helpful questions.
How did you change the layer ids of 325 and 328 to'boxes' and'classes' in the following picture?
And how did you know that the two correspond to the box and the class?

The following is the model I learned
There are 3 outputs.
The types of outputs are
My number of classes is 102.

1 is batch size 3 is channel, 80, 80 is the same height width, but I don't know what 107 is.

ps. Can the input be converted to dynamic batch when converting to onnx?

Thank you.

[glenn-jocher]

@leeyunhome it seems you have the wrong user, please refer to the follwing information:

• For how to change the layer ids of 325 and 328 to 'boxes' and 'classes' and the corresponding interpretation, you can refer to the documentation or code within the project, or reach out to the ultralytics team.
• Your question regarding the 107 output and the possibility of converting the input to dynamic batch should be addressed either by consulting the official documentation or by discussing it within the community or with the Ultralytics team.