Merge pull request #934 from feffy380/fix-minsnr-vpred-zsnr

Fix min-snr-gamma for v-prediction and ZSNR.
kohya-ss · Nov 25, 2023 · fc8649d · fc8649d
2 parents 0fb9ecf + 6b3148f
commit fc8649d
Show file tree

Hide file tree

Showing 9 changed files with 14 additions and 11 deletions.
diff --git a/fine_tune.py b/fine_tune.py
@@ -355,7 +355,7 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
                     loss = loss.mean([1, 2, 3])
 
                     if args.min_snr_gamma:
-                        loss = apply_snr_weight(loss, timesteps, noise_scheduler, 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:
                         loss = scale_v_prediction_loss_like_noise_prediction(loss, timesteps, noise_scheduler)
                     if args.debiased_estimation_loss:

diff --git a/library/custom_train_functions.py b/library/custom_train_functions.py
@@ -57,10 +57,13 @@ def enforce_zero_terminal_snr(betas):
     noise_scheduler.alphas_cumprod = alphas_cumprod
 
 
-def apply_snr_weight(loss, timesteps, noise_scheduler, gamma):
+def apply_snr_weight(loss, timesteps, noise_scheduler, gamma, v_prediction=False):
     snr = torch.stack([noise_scheduler.all_snr[t] for t in timesteps])
-    gamma_over_snr = torch.div(torch.ones_like(snr) * gamma, snr)
-    snr_weight = torch.minimum(gamma_over_snr, torch.ones_like(gamma_over_snr)).float().to(loss.device)  # from paper
+    min_snr_gamma = torch.minimum(snr, torch.full_like(snr, gamma))
+    if v_prediction:
+        snr_weight = torch.div(min_snr_gamma, snr+1).float().to(loss.device)
+    else:
+        snr_weight = torch.div(min_snr_gamma, snr).float().to(loss.device)
     loss = loss * snr_weight
     return loss
 

diff --git a/sdxl_train_control_net_lllite.py b/sdxl_train_control_net_lllite.py
@@ -460,7 +460,7 @@ def remove_model(old_ckpt_name):
                 loss = loss * loss_weights
 
                 if args.min_snr_gamma:
-                    loss = apply_snr_weight(loss, timesteps, noise_scheduler, 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:
                     loss = scale_v_prediction_loss_like_noise_prediction(loss, timesteps, noise_scheduler)
                 if args.v_pred_like_loss:

diff --git a/sdxl_train_control_net_lllite_old.py b/sdxl_train_control_net_lllite_old.py
@@ -430,7 +430,7 @@ def remove_model(old_ckpt_name):
                 loss = loss * loss_weights
 
                 if args.min_snr_gamma:
-                    loss = apply_snr_weight(loss, timesteps, noise_scheduler, 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:
                     loss = scale_v_prediction_loss_like_noise_prediction(loss, timesteps, noise_scheduler)
                 if args.v_pred_like_loss:

diff --git a/train_controlnet.py b/train_controlnet.py
@@ -449,7 +449,7 @@ def remove_model(old_ckpt_name):
                 loss = loss * loss_weights
 
                 if args.min_snr_gamma:
-                    loss = apply_snr_weight(loss, timesteps, noise_scheduler, args.min_snr_gamma)
+                    loss = apply_snr_weight(loss, timesteps, noise_scheduler, args.min_snr_gamma, args.v_parameterization)
 
                 loss = loss.mean()  # 平均なのでbatch_sizeで割る必要なし
 

diff --git a/train_db.py b/train_db.py
@@ -342,7 +342,7 @@ def train(args):
                 loss = loss * loss_weights
 
                 if args.min_snr_gamma:
-                    loss = apply_snr_weight(loss, timesteps, noise_scheduler, 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:
                     loss = scale_v_prediction_loss_like_noise_prediction(loss, timesteps, noise_scheduler)
                 if args.debiased_estimation_loss:

diff --git a/train_network.py b/train_network.py
@@ -812,7 +812,7 @@ def remove_model(old_ckpt_name):
                     loss = loss * loss_weights
 
                     if args.min_snr_gamma:
-                        loss = apply_snr_weight(loss, timesteps, noise_scheduler, 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:
                         loss = scale_v_prediction_loss_like_noise_prediction(loss, timesteps, noise_scheduler)
                     if args.v_pred_like_loss:

diff --git a/train_textual_inversion.py b/train_textual_inversion.py
@@ -578,7 +578,7 @@ def remove_model(old_ckpt_name):
                     loss = loss * loss_weights
 
                     if args.min_snr_gamma:
-                        loss = apply_snr_weight(loss, timesteps, noise_scheduler, 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:
                         loss = scale_v_prediction_loss_like_noise_prediction(loss, timesteps, noise_scheduler)
                     if args.v_pred_like_loss:

diff --git a/train_textual_inversion_XTI.py b/train_textual_inversion_XTI.py
@@ -469,7 +469,7 @@ def remove_model(old_ckpt_name):
 
                 loss = loss * loss_weights
                 if args.min_snr_gamma:
-                    loss = apply_snr_weight(loss, timesteps, noise_scheduler, 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:
                     loss = scale_v_prediction_loss_like_noise_prediction(loss, timesteps, noise_scheduler)
                 if args.debiased_estimation_loss: