The idea was that user companies (or even the hardware companies: Intel, AMD) would pay me as a contractor to implement deconvolutions, transformers, etc., as extensions to the free/open core I released (which includes very sophisticated support for all forms of dense convolution, fully-connected layers, pooling, batch norm, activations, etc.)
The free/open core supplies all the basic operators, but many users would need extensions of the core to support their particular networks (or even just software to feed their parameters into the system and get results out) and hardware companies benefit from the software's existence because they sell chips and NN-512 makes the new vector units on their chips more useful
But it didn't work out that way. NN-512 did generate offers to work on other projects (a generous offer from Intel, for example) but nothing that extends NN-512
Why? There are numerous free end-to-end deep learning tools developed by large, company-sponsored teams (Nvidia, Facebook, Google, Intel, etc.) so a specialized tool that requires integration effort is unattractive. That's my guess. So beware, if you're working along similar lines