I have compiled nanotrack onnx model that has 2 inputs: frame and template(from previous output). And i got this exception while trying to parse to har format: hailo_sdk_client.tools.parser_cli.ParserCLIException: Unexpected input_shapes at external pad layer external_pad3 (translated from /ban/corr_dw_cls/Conv), input_shapes=[[-1, 16, 16, 96], [-1, 4, 4, 96]] (type=<class ‘list’>).
I checked onnx and all inputs should have [-1, 4, 4, 96] shapes.
Updating problem:
Now layer that gives problems is Conv_50. And i dont sure Conv_56 doesnt give same problem
Maybe someone is working with Siamese arch? Or with single object trackers in general?
Hey @rprok2002 ,
I notice you’ve resolved the external_pad3 issue - great work! While I can’t identify Conv_50 and Conv_56 in the image (since the layers aren’t numbered), these operations are likely connected in your pipeline. If Conv_50 has incorrect output dimensions, it would naturally affect Conv_56 downstream.
Could you share the specific error you’re seeing with Conv_50? If it’s also a shape mismatch, we might need to add a forced reshape operation to ensure the correct dimensions. This would help resolve any cascading shape issues through your network.
I solved my problem. I removed padding from F.conv2d(x_padded, kernel, groups=batch * channel) and manually padded input.