jeffkinnison / florin

Flexible Learning-Free Reconstruction of Neural Circuits

MIT License

9 stars 2 forks source link

v0.0.2 Update #9

Closed jeffkinnison closed 5 years ago

jeffkinnison commented 5 years ago

Changelog

Enhancements

FLoRIN now uses networkx internally to create directed graphs of operations and data flow.
Dependencies may be injected by referencing other operations when setting up the pipeline.
A second MPI processing pipeline has been introduces. This pipeline only uses standard send/recv semantics and is intended for systems that do not support MPI_Comm_spawn().

Bug Fixes

FLoRIN no longer crashes after completing the final operation in the pipeline. Functionally speaking, this changes nothing, but now you won't see Python tracebacks when things succeed.
On some systems, NDNT would encounter a numpy bug when processing floating-point arrays. This was cleverly fixed by switching to integers.

Issues Addressed

6 Digraph-based Pipelines
7 Crash when ending with a Pipeline object
8 MPI - Static Scheduling and Task Execution

jeffkinnison commented 5 years ago

Unit tests are currently under development. This PR is based on real-world processing non-trivial volumes on a supercomputer.

jeffkinnison commented 5 years ago

This version has been tested on a number of volumes in a number of contexts, and is ready as a temporary update.