Loss lambda 찾기

[ThisIsIsaac]

video colorization 논문에선:
kind:weight (current value) -> weighted value

• feat: 없음 (10~15) -> ?
• perceptual: 0.001 (0.5) -> 0.0005
• contextual: 0.2 (3~2) -> 0.5
• smoothness: 5 (?) -> ?
• adversarial: 0.2 (0.5~0.4) -> 0.1
• L1 (reconstruction): 2 (0.004 ~ 0.006) -> 0.016

[ThisIsIsaac]

아직 구조가 많이 변동될 예정이라 어느정도 구조가 확립 된 후에 하는걸로.

[ThisIsIsaac]

subnet lambdas

unweighted loss value range according to source:

• L1: 0.75 ~ 0.85
• vgg: 0.055 ~ 0.06
• softmax: 10 ~ 7