VictorMylle/Kohya-ss-sd-scripts
1
0
Fork 1
mirror of https://github.com/kohya-ss/sd-scripts.git synced 2026-04-06 13:47:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix training for V-pred and ztSNR

Browse Source 
1) Updates debiased estimation loss function for V-pred.
2) Prevents now-deprecated scaling of loss if ztSNR is enabled.
This commit is contained in:
catboxanon
2024-10-21 07:34:33 -04:00
parent 012e7e63a5
commit 8fc30f8205
10 changed files with 26 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
fine_tune.py
View File

@@ -383,10 +383,10 @@ def train(args):
if args.min_snr_gamma:
loss = apply_snr_weight(loss, timesteps, noise_scheduler, args.min_snr_gamma, args.v_parameterization)
if args.scale_v_pred_loss_like_noise_pred:
if args.scale_v_pred_loss_like_noise_pred and not args.zero_terminal_snr:
loss = scale_v_prediction_loss_like_noise_prediction(loss, timesteps, noise_scheduler)
if args.debiased_estimation_loss:
loss = apply_debiased_estimation(loss, timesteps, noise_scheduler)
loss = apply_debiased_estimation(loss, timesteps, noise_scheduler, args.v_parameterization)
loss = loss.mean() # mean over batch dimension
else:

7
library/custom_train_functions.py
View File

@@ -96,10 +96,13 @@ def add_v_prediction_like_loss(loss, timesteps, noise_scheduler, v_pred_like_los
return loss
def apply_debiased_estimation(loss, timesteps, noise_scheduler):
def apply_debiased_estimation(loss, timesteps, noise_scheduler, v_prediction=False):
snr_t = torch.stack([noise_scheduler.all_snr[t] for t in timesteps]) # batch_size
snr_t = torch.minimum(snr_t, torch.ones_like(snr_t) * 1000) # if timestep is 0, snr_t is inf, so limit it to 1000
weight = 1 / torch.sqrt(snr_t)
if v_prediction:
weight = 1 / (snr_t + 1)
else:
weight = 1 / torch.sqrt(snr_t)
loss = weight * loss
return loss

5
library/train_util.py
View File

@@ -3731,6 +3731,11 @@ def verify_training_args(args: argparse.Namespace):
raise ValueError(
"scale_v_pred_loss_like_noise_pred can be enabled only with v_parameterization / scale_v_pred_loss_like_noise_predはv_parameterizationが有効なときのみ有効にできます"
)
if args.scale_v_pred_loss_like_noise_pred and args.zero_terminal_snr:
raise ValueError(
"zero_terminal_snr enabled. scale_v_pred_loss_like_noise_pred will not be used / zero_terminal_snrが有効です。scale_v_pred_loss_like_noise_predは使用されません"
)
if args.v_pred_like_loss and args.v_parameterization:
raise ValueError(

4
sdxl_train.py
View File

@@ -725,12 +725,12 @@ def train(args):
if args.min_snr_gamma:
loss = apply_snr_weight(loss, timesteps, noise_scheduler, args.min_snr_gamma, args.v_parameterization)
if args.scale_v_pred_loss_like_noise_pred:
if args.scale_v_pred_loss_like_noise_pred and not args.zero_terminal_snr:
loss = scale_v_prediction_loss_like_noise_prediction(loss, timesteps, noise_scheduler)
if args.v_pred_like_loss:
loss = add_v_prediction_like_loss(loss, timesteps, noise_scheduler, args.v_pred_like_loss)
if args.debiased_estimation_loss:
loss = apply_debiased_estimation(loss, timesteps, noise_scheduler)
loss = apply_debiased_estimation(loss, timesteps, noise_scheduler, args.v_parameterization)
loss = loss.mean() # mean over batch dimension
else:

4
sdxl_train_control_net_lllite.py
View File

@@ -474,12 +474,12 @@ def train(args):
if args.min_snr_gamma:
loss = apply_snr_weight(loss, timesteps, noise_scheduler, args.min_snr_gamma, args.v_parameterization)
if args.scale_v_pred_loss_like_noise_pred:
if args.scale_v_pred_loss_like_noise_pred and not args.zero_terminal_snr:
loss = scale_v_prediction_loss_like_noise_prediction(loss, timesteps, noise_scheduler)
if args.v_pred_like_loss:
loss = add_v_prediction_like_loss(loss, timesteps, noise_scheduler, args.v_pred_like_loss)
if args.debiased_estimation_loss:
loss = apply_debiased_estimation(loss, timesteps, noise_scheduler)
loss = apply_debiased_estimation(loss, timesteps, noise_scheduler, args.v_parameterization)
loss = loss.mean() # 平均なのでbatch_sizeで割る必要なし

4
sdxl_train_control_net_lllite_old.py
View File

@@ -434,12 +434,12 @@ def train(args):
if args.min_snr_gamma:
loss = apply_snr_weight(loss, timesteps, noise_scheduler, args.min_snr_gamma, args.v_parameterization)
if args.scale_v_pred_loss_like_noise_pred:
if args.scale_v_pred_loss_like_noise_pred and not args.zero_terminal_snr:
loss = scale_v_prediction_loss_like_noise_prediction(loss, timesteps, noise_scheduler)
if args.v_pred_like_loss:
loss = add_v_prediction_like_loss(loss, timesteps, noise_scheduler, args.v_pred_like_loss)
if args.debiased_estimation_loss:
loss = apply_debiased_estimation(loss, timesteps, noise_scheduler)
loss = apply_debiased_estimation(loss, timesteps, noise_scheduler, args.v_parameterization)
loss = loss.mean() # 平均なのでbatch_sizeで割る必要なし

4
train_db.py
View File

@@ -370,10 +370,10 @@ def train(args):
if args.min_snr_gamma:
loss = apply_snr_weight(loss, timesteps, noise_scheduler, args.min_snr_gamma, args.v_parameterization)
if args.scale_v_pred_loss_like_noise_pred:
if args.scale_v_pred_loss_like_noise_pred and not args.zero_terminal_snr:
loss = scale_v_prediction_loss_like_noise_prediction(loss, timesteps, noise_scheduler)
if args.debiased_estimation_loss:
loss = apply_debiased_estimation(loss, timesteps, noise_scheduler)
loss = apply_debiased_estimation(loss, timesteps, noise_scheduler, args.v_parameterization)
loss = loss.mean() # 平均なのでbatch_sizeで割る必要なし

4
train_network.py
View File

@@ -993,12 +993,12 @@ class NetworkTrainer:
if args.min_snr_gamma:
loss = apply_snr_weight(loss, timesteps, noise_scheduler, args.min_snr_gamma, args.v_parameterization)
if args.scale_v_pred_loss_like_noise_pred:
if args.scale_v_pred_loss_like_noise_pred and not args.zero_terminal_snr:
loss = scale_v_prediction_loss_like_noise_prediction(loss, timesteps, noise_scheduler)
if args.v_pred_like_loss:
loss = add_v_prediction_like_loss(loss, timesteps, noise_scheduler, args.v_pred_like_loss)
if args.debiased_estimation_loss:
loss = apply_debiased_estimation(loss, timesteps, noise_scheduler)
loss = apply_debiased_estimation(loss, timesteps, noise_scheduler, args.v_parameterization)
loss = loss.mean() # 平均なのでbatch_sizeで割る必要なし

4
train_textual_inversion.py
View File

@@ -598,12 +598,12 @@ class TextualInversionTrainer:
if args.min_snr_gamma:
loss = apply_snr_weight(loss, timesteps, noise_scheduler, args.min_snr_gamma, args.v_parameterization)
if args.scale_v_pred_loss_like_noise_pred:
if args.scale_v_pred_loss_like_noise_pred and not args.zero_terminal_snr:
loss = scale_v_prediction_loss_like_noise_prediction(loss, timesteps, noise_scheduler)
if args.v_pred_like_loss:
loss = add_v_prediction_like_loss(loss, timesteps, noise_scheduler, args.v_pred_like_loss)
if args.debiased_estimation_loss:
loss = apply_debiased_estimation(loss, timesteps, noise_scheduler)
loss = apply_debiased_estimation(loss, timesteps, noise_scheduler, args.v_parameterization)
loss = loss.mean() # 平均なのでbatch_sizeで割る必要なし

4
train_textual_inversion_XTI.py
View File

@@ -483,10 +483,10 @@ def train(args):
loss = loss * loss_weights
if args.min_snr_gamma:
loss = apply_snr_weight(loss, timesteps, noise_scheduler, args.min_snr_gamma, args.v_parameterization)
if args.scale_v_pred_loss_like_noise_pred:
if args.scale_v_pred_loss_like_noise_pred and not args.zero_terminal_snr:
loss = scale_v_prediction_loss_like_noise_prediction(loss, timesteps, noise_scheduler)
if args.debiased_estimation_loss:
loss = apply_debiased_estimation(loss, timesteps, noise_scheduler)
loss = apply_debiased_estimation(loss, timesteps, noise_scheduler, args.v_parameterization)
loss = loss.mean() # 平均なのでbatch_sizeで割る必要なし