ONNX-Net: Towards Universal Representations and Instant Performance Prediction for Neural Architectures