Writing on neural network verification, tools and practice.https://zhongkuima.github.io/https://zhongkuima.github.io/blogs/2026/shapeonnx/https://zhongkuima.github.io/blogs/2026/shapeonnx/A dual-track shape inference tool that resolves ONNX's dynamic shapes to concrete static values for neural network verification workflows.Thu, 01 Jan 2026 00:00:00 GMTONNXShape InferenceStatic Analysishttps://zhongkuima.github.io/blogs/2026/slimonnx/https://zhongkuima.github.io/blogs/2026/slimonnx/A pure Python toolkit for optimizing ONNX models specifically for verification workflows, validated on the VNN-COMP 2024 benchmark suites.Thu, 01 Jan 2026 00:00:00 GMTONNX OptimizationGraph SimplificationVerificationhttps://zhongkuima.github.io/blogs/2026/torchonnx/https://zhongkuima.github.io/blogs/2026/torchonnx/A pure Python compiler that converts ONNX models to native PyTorch code through a 6-stage pipeline, achieving 100% success on VNN-COMP 2024 benchmarks.Thu, 01 Jan 2026 00:00:00 GMTONNXPyTorchCompiler DesignModel Conversionhttps://zhongkuima.github.io/blogs/2026/torchvnnlib/https://zhongkuima.github.io/blogs/2026/torchvnnlib/A tool for converting VNN-LIB verification properties to binary formats, achieving 10-100x faster loading with a two-tier processing architecture.Thu, 01 Jan 2026 00:00:00 GMTNeural Network VerificationVNN-LIBPerformance Optimizationhttps://zhongkuima.github.io/blogs/2026/wraact/https://zhongkuima.github.io/blogs/2026/wraact/A unified framework for tight convex hull approximation of neural network activation functions, supporting 7 core activation types with sub-millisecond performance.Thu, 01 Jan 2026 00:00:00 GMTPolytope ApproximationConvex HullActivation Functionshttps://zhongkuima.github.io/blogs/2026/wralu/https://zhongkuima.github.io/blogs/2026/wralu/A fast and precise approach to over-approximating the convex hull of the ReLU function, reporting 10x-10^6x runtime improvements and up to 50% fewer constraints in the evaluated benchmarks.Thu, 01 Jan 2026 00:00:00 GMTReLUConvex HullVerificationPOPL