แนวคิดหลัก
OT-based GANs benefit from strictly convex functions and cost functions to enhance stability and prevent mode collapse.
บทคัดย่อ
Optimal Transport (OT) theory is utilized in generative modeling, with recent advancements focusing on OT-based generative models. These models face challenges like unstable training processes and limited expressivity. UOTM introduces a semi-dual form of Unbalanced Optimal Transport (UOT) problem, showing promising outcomes but lacking τ-robustness. A novel method, UOTM with Scheduled Divergence (UOTM-SD), gradually adjusts the divergence term to improve performance while addressing τ-sensitivity. The equi-Lipschitz continuity of UOTM potential contributes to stable convergence during training.
สถิติ
Lvϕ,Tθ = inf 1 |X| P x∈X g1 (-c(x, ˆy) + vϕ(ˆy)) + 1 |Y| P y∈Y g2(-vϕ(y))
Lv = 1 |X| P x∈X g3((c(x, Tθ(x)) - vϕ(Tθ(x))))
FID score of UOTM-SD: 2.51 on CIFAR-10 and 5.99 on CelebA-HQ-256.