Some parts of scientific programs often can be considered to be directed acyclic graphs which edges are matrices or tensors and which nodes are operations of linear algebra. During the run of the program often only some of the inputs are changed. What libraries exist that notices which inputs changed and only recalculate those parts of the DAG that are effected by the change and do not recalculate the parts of the DAG that are not effected?
The DAGs are rather short but recalculated many times, thus tensorflow would not be efficient - the overhead of transforming the numpy arrays to tensorflow (and esp. bringing the arrays from RAM into the GPU, but even the othe CPU structures) is too large.
question from:
https://stackoverflow.com/questions/65874364/what-libraries-exist-for-reactive-dags-for-python 与恶龙缠斗过久,自身亦成为恶龙；凝视深渊过久,深渊将回以凝视…
