junyanz / CycleGAN

Software that can generate photos from paintings, turn horses into zebras, perform style transfer, and more.
Other
12.41k stars 1.95k forks source link

Training with limited/less data #121

Closed AadSah closed 5 years ago

AadSah commented 5 years ago

@junyanz Sir, Is there any trick or better way to train a CycleGAN if I have very fewer data from one domain but sufficient data from the other domain?

junyanz commented 5 years ago

One idea is to (1) train an auto-encoder for the domain with sufficient data. (2) share the generator across two domains. See some recent work for more details [1], [2].