by LeviFiction » Thu Feb 14, 2019 12:49 am
Unnecessary response to rhetorical question: *shrug* I dunno. xD
I can probably assume that they have a very good reason for using the version of tensorflow that they do. It does support newer cpu commands that older versions of tensorflow do not. And using this older version may get the plugin working but might have unforeseen drawbacks. It's hard to say really. The idea for going back to an older version of tensorflow came from stack-overflow questions that mentioned the same error LindaLB and iraisok reported. And that recommendation was to go back to version 1.6 not 1.5. We found out 1.5 worked for iraisok and not 1.6 by trying several versions of tensorflow with including the far newer 1.12. Only version 1.5 worked. No idea why.
Technically, if they didn't mind slowing things down, they could use tensorflow.js and avoid all of this nonsense. The software they use to present the UI is called Electron and it uses javascript (node.js) and HTML(chromium) to run the application. They would need to convert their trained models to the open format that tensorflow.js uses. But they wouldn't have to use separate compiled python programs to get the job done. Just do it all inside electron using javascript.