The main change here is avoiding storing simulator state for signals targeted by set operations (since that state will never be read).
I also added some more graph simplification functions. One (remove_constant_copies), I had previously disabled because it could change which signals are stored by save_params/load_params. But thinking about this more, that doesn't seem worth disabling it for. As long as the user is consistently using (or not using) remove_constant_copies in both networks, then save/load params will work fine. So better to have it enabled by default (users can still disable it if they run into problems).
The second one remove_reset_incs, I had previously completely removed because in general it seemed to make things worse. But I realized that this was largely to do with the kinds of networks we often build in Nengo, which have a lot of horizontal parallelism. In more deep learning style networks, where you just have a big long chain of objects, remove_reset_incs can still be helpful. And then with some further testing, I didn't really see any significant performance decreases when enabling it by default, so I just went with that.
The main change here is avoiding storing simulator state for signals targeted by
setoperations (since that state will never be read).I also added some more graph simplification functions. One (
remove_constant_copies), I had previously disabled because it could change which signals are stored bysave_params/load_params. But thinking about this more, that doesn't seem worth disabling it for. As long as the user is consistently using (or not using)remove_constant_copiesin both networks, then save/load params will work fine. So better to have it enabled by default (users can still disable it if they run into problems).The second one
remove_reset_incs, I had previously completely removed because in general it seemed to make things worse. But I realized that this was largely to do with the kinds of networks we often build in Nengo, which have a lot of horizontal parallelism. In more deep learning style networks, where you just have a big long chain of objects,remove_reset_incscan still be helpful. And then with some further testing, I didn't really see any significant performance decreases when enabling it by default, so I just went with that.