Same thing in deep learning. I already had to dump half of my code because of theano, as well as api changes in keras and tensorflow that made it a pain to load

Same thing in deep learning. I already had to dump half of my code because of theano, as well as api changes in keras and tensorflow that made it a pain to load some of my old models in newer versions of the frameworks. Now I'm rewriting a bunch of it in pytorch while anxiously waiting for 0.4 to come out (no idea when) and break all of my models again.