DeepTrainer roadmap changes

More than a year passed since the last productive post, since then I have got busy with a daytime job. In the meantime I have been thinking a lot where I could define the main selling point of this project if I wanted to make it to generate income for me. I stopped making the source code public, but not because I could sell it to anyone, I do not think I could sell this library for much money, especially not in the days of such excellent free tools as TensorFlow.

The selling point of this project is elsewhere, where niche, parallel, highly scalable algorithm development experience is required, not just pre-compiled libraries or copy-pasted source code. If you need someone who could implement and/or maintain similar algorithms for your data center, optimizing for a large number of Xeon-Phi, ARM, Cuda or bespoke nodes to do image-recognition or other types of learning tasks – just contact me.

I am also happy to answer any questions related to the blog and the libraries, just please make sure you understand what all of this is about. The last time someone contacted me because of my project they were asking for my contribution to a startup project, and up until we were deep in some discussions on the phone they did not even realise that I am implementing my own algorithms from scratch, and they were just looking for someone who knows tensorflow/python scripting. Or at least I guess that is what they were looking for.

In the meantime I am going to continue where I last left off:

  • moving to platform-independence with the help of QT
  • installable demo programs both on Windows and Linux OSs
  • a new layer-granularity algorithm design to allow for even more algorithms types
  • adding CNN convolutional layers to enable image and shape recognition
  • implementing a highly scalable microservice architecture

In the meantime if you feel like to supporting my work you can find the donate button below.

Leave a Reply