Closed erikbern closed 6 years ago
Get rid of all the minibatch stuff. This simplifies the code a lot and seems to improve convergence and also make it faster.
Only downside is we can't do fun stuff like dropout but we're not doing that right now anyway.
Coverage decreased (-0.7%) to 92.717% when pulling 7811659ce5590842ed7859cef49f512ea0830ae4 on weights into e9616bec25f14ce5cc3fad483d0382b03b06f9c8 on master.
Get rid of all the minibatch stuff. This simplifies the code a lot and seems to improve convergence and also make it faster.
Only downside is we can't do fun stuff like dropout but we're not doing that right now anyway.