Add option to use Scheduled Huber Loss in all training pipelines to improve resilience to data corruption #1228
add huber loss and huber_c compute to train_util
7e346b08
add reduction modes
9a7112ac
add huber_c retrieval from timestep getter
7da9bcf6
move get timesteps and huber to own function
81e26183
add conditional loss to all training scripts
029fcf87
add cond loss to train network
117a207b
add (scheduled) huber_loss to args
b1085b95
fixup twice timesteps getting
8abc1128
PHL-schedule should depend on noise scheduler's num timesteps
a58f290b
kabachuha
marked this pull request as draft 2 years ago
*2 multiplier to huber loss cause of 1/2 a^2 conv.
c6495def
add option for smooth l1 (huber / delta)
dd22958c
unify huber scheduling
19a834c3
add snr huber scheduler
47fb1a68
kabachuha
marked this pull request as ready for review 2 years ago
kohya-ss
changed the base branch from
main
to
dev
2 years ago
Merge branch 'dev' into main
8ab8e4a5
kohya-ss
merged
90b18795
into dev 2 years ago
Assignees
No one assigned
Login to write a write a comment.
Login via GitHub