1 个仓库
Tools that identify specific sequences of operations in a graph and replace them with optimized equivalents.
Distinct from Pattern-Based Replacements: Targets computational graphs of neural networks rather than AST nodes of source code.
Explore 1 awesome GitHub repository matching software engineering & architecture · Graph Pattern Rewriters. Refine with filters or upvote what's useful.
onnxsim 是一个深度学习图优化器和模型简化器,旨在降低 ONNX 计算图的复杂度。它充当模型压缩器,用简化的常量输出替换复杂的算子序列,以减少操作开销。 该项目通过常量折叠推理实现简化,将常量算子的子图替换为预计算的常量张量。它利用基于模式的图重写和静态计算图分析来识别并删除冗余节点或不可达操作。 该工具涵盖了广泛的模型优化功能,包括算子冗余消除以及删除不必要的 reshape 或 identity 节点。这些过程简化了执行流程并减少了模型的内存占用。
Uses pattern-based rewriting to replace operator sequences with simplified equivalent structures.