We have improved the implementation of MixStyle to make it more flexible.
Recall that MixStyle has two versions: random mixing and cross-domain mixing. The former randomly shuffles the batch dimension while the latter mixes the 1st half in a batch with the 2nd half.
After merging MixStyle2 to MixStyle, the two versions are now managed by a new variable called self.mix, which takes as input either random or crossdomain that correspond to the two versions respectively. This variable can be set during initialization, e.g., self.mixstyle = MixStyle(mix='random'). It can also be changed on-the-fly. For instance, say you wanna apply random mixing at current step, simply do model.apply(random_mixstyle), or model.apply(crossdomain_mixstyle) if you prefer the cross-domain mixing manner.
We have also added new context managers to manage mixstyle in the forward pass. Say your model has MixStyle layers which were initially activated and you would like to deactivate them at a certain time, you can do
But note that the change in self.mix during run_with_mixstyle is permanent unless you manually use model.apply(random_mixstyle) or model.apply(crossdomain_mixstyle) to modify the variable.
We have improved the implementation of MixStyle to make it more flexible.
Recall that MixStyle has two versions: random mixing and cross-domain mixing. The former randomly shuffles the batch dimension while the latter mixes the 1st half in a batch with the 2nd half.
After merging
MixStyle2toMixStyle, the two versions are now managed by a new variable calledself.mix, which takes as input eitherrandomorcrossdomainthat correspond to the two versions respectively. This variable can be set during initialization, e.g.,self.mixstyle = MixStyle(mix='random'). It can also be changed on-the-fly. For instance, say you wanna apply random mixing at current step, simply domodel.apply(random_mixstyle), ormodel.apply(crossdomain_mixstyle)if you prefer the cross-domain mixing manner.We have also added new context managers to manage mixstyle in the forward pass. Say your model has MixStyle layers which were initially activated and you would like to deactivate them at a certain time, you can do
Otherwise if you want to use MixStyle layers which were initially deactivated, you can do
You can also change
self.mixwhile usingrun_with_mixstyle, e.g.But note that the change in
self.mixduringrun_with_mixstyleis permanent unless you manually usemodel.apply(random_mixstyle)ormodel.apply(crossdomain_mixstyle)to modify the variable.