Closed collinarnett closed 4 years ago
75c402dde337061f44189a1b32328763dcfafd59 added some new changes to better reflect paper for better training stability.
I've tried to implement changes to decrease loss of the generator but I can't seem to hack it without changing the entire architecture. One sided label smoothing definitely helps but is not perfect and introducing noice every 20 steps does not improve model learning.
Going to leave this now since the paper is now fully replicated.
Currently there are a few hyper parameters that aren't mentioned in the paper that need to be optimized: