ONNX spec: "kernel_shape : list of ints The shape of the convolution kernel. If not present, should be inferred from input W." We were taking the number of input/output feature maps into account, which we realized we should not.