Conv.W key not found for yolo onnx models

[IanButterworth]

I'm trying to convert the tiny-yolo onnx models generated in this blog post, which are hosted here, and I'm getting the following errors:

>using Flux, ONNX
>ONNX.load_model("20180709 TinyYoloV3.onnx")
KeyError: key "Conv.W" not found

>ONNX.load_model("Tiny-YOLOv2.onnx")
KeyError: key "convolution.W" not found

Is this a peculiarity of these ONNX models, or an unsupported ONNX standard? I ask because I know that the tiny-yolov2 onnx model here does convert successfully (but is not accurate due to the asymmetrical padding issue with yolov2, hence why I'm trying to get yolov3 up and running)

[ayush1999]

@ianshmean There's an error in the onnx model itself: See onnx/models#60.

[IanButterworth]

Ah ok. Is a workaround possible in the interim to just ignore keys that aren’t initialized?

[IanButterworth]

Seems that it's an issue with <1.2 onnx, given the 1.0 version here is the only one that causes this issue. ONNX seems to have moved forward a lot since 1.0 so perhaps the best plan would be to generate newer ONNX models

[ayush1999]

You could try downloading the pretrained model in some other framework (PyTorch/TensorFlow) and use the newer version of onnx exporters to generate the onnx file, and load it in using this package. (Would also be a great addition for the onnx/models repo)