Closed siboehm closed 2 years ago
Ref #66
The non-linear disentanglement loss is replaced by a MLP with 2 hidden layers. During the hparam sweeps most adversarial classifiers had dimensions 64 with 2-3 hidden layers, so now the model we use during eval is similar.
I also added a new Yaml file for #69
Ref #66
The non-linear disentanglement loss is replaced by a MLP with 2 hidden layers. During the hparam sweeps most adversarial classifiers had dimensions 64 with 2-3 hidden layers, so now the model we use during eval is similar.