This is just a personal question. I was wondering if it is possible to use or port the TensorFlow library to the CE Toolchain for small-scale neural CNN networks. I know that there is a C port of the library available here.

I know that training neural networks is time-consuming and requires a lot of computational power. I was planning on doing training on a computer and porting the network weights over to the calculator, where there will be a small version of the model where the weights will be transferred.

Does this sound possible, and will there be any limitations I might face along the way?

I did this back in March: https://z80.me/blog/calculator-ai-part-1/ , including implementing training on the calculator as a proof-of-concept but then training on a computer and transferring the weights to the calculator. Significant limitations exist, especially with the memory a network takes and the speed of floating-point operations on the ez80. I have plans sketched out for some optimizations, such as quantized weights and use of integral types, but there's not terribly much practical point.

You can definitely do tensor operations on the CE - I believe Kerm was able to get a handwriting recognition model running on it a while back, though I can't seem to find the thread for that now.

I don't have a whole lot of experience with machine learning, but I tried TensorFlow a few years ago and I was under the impression that it's a "kitchen sink" framework of sorts for ML applications, rather than a Unix philosophy program that does a single thing well. If you're just trying to do inference on a single type of model, it would probably be easier to implement the operations you need from scratch rather than trying to port all of the necessary surrounding infrastructure from TensorFlow to the CE.