You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
keicoon15 edited this page Aug 21, 2018
·
1 revision
Deep convolutional GAN은 기존의 GAN 모델의 네트워크 구조에 dense layer 대신 convolutional layer를 사용한 특징을 가지고 있다.
architecture에서는 별도의 pooling 이나 unspooling 레이어를 사용하지 않고 Generator의 spatial dimension을 증가시킬 필요가 있을 때에만 transposed convolution를 사용한다.
DCGAN은 제한된 도메인의 이미지에 대해서 매우 높은 퀄리티의 이미지를 생성해주고 이것은 DCGAN 모델이 latent code를 유의미하게 사용하는 방법을 배웠다는 것을 말해준다. 그리고 latent space 에서 간단한 사칙연산을 통해 이미지의 semantic한 특성들을 operation 할 수도 있다.