taming-transformers
taming-transformers copied to clipboard
Deciding when to initiate discriminator loss to kick in
The argument disc_start
of configuration yaml file is numbers of steps when discriminator loss kicks in during the training.
The significance of the question is based on issue #93 where loss oscillates after discriminator loss gets involved in training process. It seems like loss value often ticks below zero due to discriminator loss also.
- First question is, What will be the recommended steps(or epochs) for discriminator iteration starting point(disc_start)?
- Second question is, shouldn't it be better to set as relative percentage of the total steps for training rather than absolute value of steps?
Below is part of configs/custom_vqgan.yaml file with default values, FIY
lossconfig:
target: taming.modules.losses.vqperceptual.VQLPIPSWithDiscriminator
params:
disc_conditional: False
disc_in_channels: 3
disc_start: 10000 # 10000 steps
disc_weight: 0.8
codebook_weight: 1.0
bump does somebody have an answer to this?