Question concerning exporting to ONNX

[gillmac13]

@glenn-jocher

I have been using your repo time and again with excellent results for 1-class detection with yolov3-spp. I am now working on different ways to deploy some models which need a .onnx format.
However, when applying the suggested conversion method to the letter, I get an unexpected output ...

For example, if I download a yolov3-spp.weights model (80 classes) from https://pjreddie.com/darknet/yolo/ , and:

• change to ONNX_EXPORT = True in models.py
• run the command: python3 detect.py --cfg cfg/yolov3-spp.cfg --weights weights/yolov3-spp.weights

I get an export.onnx file in return with no warning or error.
Now if I inspect this export.onnx with netron, I find that the output of the model is 2 tensors, 1 of shape 10647x80 and 1 of shape 10647x1. I was under the impression that an 80-class model would produce 85 columns, not 84.

Using the code snippet:
model_onnx = onnxruntime.InferenceSession("weights/export.onnx") ort_inputs = {model_onnx.get_inputs()[0].name: img_t.numpy()} ort_outs = model_onnx.run(None, ort_inputs) x_out = ort_outs[0] print(x_out.shape)

confirms what I get in terms of output format.
And applying this procedure to my 1-class models, I get 4+1 columns instead of the expected 6, such as in the "pred" variable in "detect.py".

My questions are simple: is what I get in terms of output normal, in which case how should I interpret the results, or is something wrong?

[gillmac13]

Correction from above, I meant 1 tensor of shape 10647x80 et 1 of shape 10647x4

[glenn-jocher]

@gillmac13 yes this ONNX output is correct. These are the 4 box coordinates and the 80 class confidences, which have been multiplied by the objectness confidence.

See #653 (comment) for a stricter conversion which retains the 85 outputs, though be aware that the export pipeline is prone to failure on many pytorch-onnx version combinations.

[gillmac13]

@glenn-jocher

Thanks for your prompt reply.
And that clears up the format aspect of things.
Now if I dig up the code, I see that in order to perform NMS on the model output, the 2 confidences are needed. With them merged, is there another way to compute NMS ?

[glenn-jocher]

@gillmac13 NMS is not supplied in the ONNX model, nor in the PyTorch nor darknet models. NMS operations are executed after the model runs inference.