VictorMylle/Kohya-ss-sd-scripts
1
0
Fork 1
mirror of https://github.com/kohya-ss/sd-scripts.git synced 2026-04-08 22:35:09 +00:00
Code Issues Packages Projects Releases Wiki Activity

Implement pseudo Huber loss for Flux and SD3

Browse Source
This commit is contained in:
recris
2024-11-27 18:11:51 +00:00
parent 2a61fc0784
commit 420a180d93
15 changed files with 76 additions and 61 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
sdxl_train.py
View File

@@ -695,7 +695,7 @@ def train(args):
# Sample noise, sample a random timestep for each image, and add noise to the latents,
# with noise offset and/or multires noise if specified
noise, noisy_latents, timesteps, huber_c = train_util.get_noise_noisy_latents_and_timesteps(
noise, noisy_latents, timesteps = train_util.get_noise_noisy_latents_and_timesteps(
args, noise_scheduler, latents
)
@@ -720,7 +720,7 @@ def train(args):
):
# do not mean over batch dimension for snr weight or scale v-pred loss
loss = train_util.conditional_loss(
noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c
args, noise_pred.float(), target.float(), timesteps, "none", noise_scheduler
)
if args.masked_loss or ("alpha_masks" in batch and batch["alpha_masks"] is not None):
loss = apply_masked_loss(loss, batch)
@@ -738,7 +738,7 @@ def train(args):
loss = loss.mean() # mean over batch dimension
else:
loss = train_util.conditional_loss(
noise_pred.float(), target.float(), reduction="mean", loss_type=args.loss_type, huber_c=huber_c
args, noise_pred.float(), target.float(), timesteps, "none", noise_scheduler
)
accelerator.backward(loss)