I haven't blogged here in a long time. Here's some info about a project I've been working on.
Recently I've been toying with Google's Tensorflow python library. I've written an op for the cpu and the gpu. The op, of course, is a modified version of the Dijkstra algorithm. The funny thing is that the cpu version works faster than the gpu version. There seems to be a long overhead time, something like 7 seconds, before a gpu op begins to work. There is also a size restriction. My op works on grids. On the cpu version I can use a large grid. I've tried 480x480. On the gpu the op fails if the grid is larger than 70x70. What do I come away with? Well, I enjoyed doing the gpu version, but the cpu version ends up being superior.
So, the code is here: https://github.com/radiodee1/awesome-tf
I also use pygame to construct my gui for testing with real png images.