The scheme between torch.nn.functional.Pad and onnx's pad is different if the pads receive 8-dimensional padding.
Given input with size: (1,3,10,10) and pads=(1,1,2,2,3,3,4,4), F.Pad will result in size (9,9,14,12) (as declared in your test scripts). However, onnx will output result with size: (5,7,16,16) following their documentation.

Therefore, the pads parameter loaded from onnx_model.graph should be transformed to the PyTorch version so the padding size is correct.

Unfortunately, I find that the pads parameter will be placed in onnx_model.graph.initializer instead of the node's parameter, so a simple preprocess of Pad nodes' parameter is not feasible :(.

So I have to add an additional branch (which is ugly...) when loading the initializer parameter: if the targeting node is Pad we will check if the pads parameter needs to be preprocessed.

I write a program to exhibit this bug:
Through this code snippet, we can see that the correct output shape should be: (batch_size, 226, 226, 3) but ONNX2PyTorch will output (batch_size+1, 225, 225, 3)

Current fix can pass all existing tests.