Preserving model outputs during optimization
Folding BatchNorm into Gemm ignored Gemm’s beta coefficient, changing the output of optimized ONNX models. The fix includes beta in the fused bias and adds parameterized regression coverage.
Selected fixes in numerical computing, distributed training, and Windows compatibility. Each example links to its upstream pull request.
Folding BatchNorm into Gemm ignored Gemm’s beta coefficient, changing the output of optimized ONNX models. The fix includes beta in the fused bias and adds parameterized regression coverage.
Renaming a file to a different case made Windows report a deletion that never happened, so anything watching the folder saw a file vanish and come back. The false event is coalesced away now.
Whenever two values tied, the same paired data gave a different answer depending on the order the rows arrived in. Ties now share an average rank. Three further fixes followed, including line endings that broke linting on a fresh Windows clone.
Three fixes to a film-negative tool. The Windows updater broke on any profile path containing a non-ASCII character, script generation assumed POSIX paths, and the tests read source files in whatever codepage the machine happened to use.