Min-SNR Weighting Strategy

fine_tune.py

@@ -19,7 +19,8 @@
     ConfigSanitizer,
     BlueprintGenerator,
 )
-
+import library.custom_train_functions as custom_train_functions
+from library.custom_train_functions import apply_snr_weight 
 
 def collate_fn(examples):
     return examples[0]
@@ -304,6 +305,9 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
 
                 loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float(), reduction="mean")
 
+                if args.min_snr_gamma:
+                  loss = apply_snr_weight(loss, latents, noisy_latents, args.min_snr_gamma)
+
                 accelerator.backward(loss)
                 if accelerator.sync_gradients and args.max_grad_norm != 0.0:
                     params_to_clip = []
@@ -396,6 +400,8 @@ def setup_parser() -> argparse.ArgumentParser:
     train_util.add_sd_saving_arguments(parser)
     train_util.add_optimizer_arguments(parser)
     config_util.add_config_arguments(parser)
+    custom_train_functions.add_custom_train_arguments(parser)
+
 
     parser.add_argument("--diffusers_xformers", action="store_true", help="use xformers by diffusers / Diffusersでxformersを使用する")
     parser.add_argument("--train_text_encoder", action="store_true", help="train text encoder / text encoderも学習する")

library/custom_train_functions.py

@@ -0,0 +1,17 @@
+import torch
+import argparse
+
+def apply_snr_weight(loss, latents, noisy_latents, gamma):
+  sigma = torch.sub(noisy_latents, latents) #find noise as applied by scheduler
+  zeros = torch.zeros_like(sigma) 
+  alpha_mean_sq = torch.nn.functional.mse_loss(latents.float(), zeros.float(), reduction="none").mean([1, 2, 3]) #trick to get Mean Square/Second Moment
+  sigma_mean_sq = torch.nn.functional.mse_loss(sigma.float(), zeros.float(), reduction="none").mean([1, 2, 3]) #trick to get Mean Square/Second Moment
+  snr = torch.div(alpha_mean_sq,sigma_mean_sq) #Signal to Noise Ratio = ratio of Mean Squares
+  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() #from paper
+  loss = loss * snr_weight
+  print(snr_weight)
+  return loss
+
+def add_custom_train_arguments(parser: argparse.ArgumentParser):
+  parser.add_argument("--min_snr_gamma", type=float, default=0, help="gamma for reducing the weight of high loss timesteps. Lower numbers have stronger effect. 5 is recommended by paper.")

library/train_util.py

@@ -1963,7 +1963,7 @@ def add_training_arguments(parser: argparse.ArgumentParser, support_dreambooth:
         parser.add_argument(
             "--prior_loss_weight", type=float, default=1.0, help="loss weight for regularization images / 正則化画像のlossの重み"
         )
-    parser.add_argument("--min_snr_gamma", type=float, default=0, help="gamma for reducing the weight of high loss timesteps. Lower numbers have stronger effect. 5 is recommended by paper.")
+
 
 def verify_training_args(args: argparse.Namespace):
     if args.v_parameterization and not args.v2:

train_db.py

@@ -21,7 +21,8 @@
     ConfigSanitizer,
     BlueprintGenerator,
 )
-
+import library.custom_train_functions as custom_train_functions
+from library.custom_train_functions import apply_snr_weight 
 
 def collate_fn(examples):
     return examples[0]
@@ -291,6 +292,9 @@ def train(args):
                 loss_weights = batch["loss_weights"]  # 各sampleごとのweight
                 loss = loss * loss_weights
 
+                if args.min_snr_gamma:
+                  loss = apply_snr_weight(loss, latents, noisy_latents, args.min_snr_gamma)
+
                 loss = loss.mean()  # 平均なのでbatch_sizeで割る必要なし
 
                 accelerator.backward(loss)
@@ -390,6 +394,7 @@ def setup_parser() -> argparse.ArgumentParser:
     train_util.add_sd_saving_arguments(parser)
     train_util.add_optimizer_arguments(parser)
     config_util.add_config_arguments(parser)
+    custom_train_functions.add_custom_train_arguments(parser)
 
     parser.add_argument(
         "--no_token_padding",

train_network.py

@@ -23,7 +23,8 @@
     ConfigSanitizer,
     BlueprintGenerator,
 )
-
+import library.custom_train_functions as custom_train_functions
+from library.custom_train_functions import apply_snr_weight 
 
 def collate_fn(examples):
     return examples[0]
@@ -548,16 +549,9 @@ def train(args):
 
                 loss_weights = batch["loss_weights"]  # 各sampleごとのweight
                 loss = loss * loss_weights
-                gamma = args.min_snr_gamma
-                if gamma:
-                  sigma = torch.sub(noisy_latents, latents) #find noise as applied
-                  zeros = torch.zeros_like(sigma) 
-                  alpha_mean_sq = torch.nn.functional.mse_loss(latents.float(), zeros.float(), reduction="none").mean([1, 2, 3]) #trick to get Mean Square
-                  sigma_mean_sq = torch.nn.functional.mse_loss(sigma.float(), zeros.float(), reduction="none").mean([1, 2, 3]) #trick to get Mean Square
-                  snr = torch.div(alpha_mean_sq,sigma_mean_sq) #Signal to Noise Ratio = ratio of Mean Squares
-                  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() #from paper
-                  loss = loss * snr_weight
+
+                if args.min_snr_gamma:
+                  loss = apply_snr_weight(loss, latents, noisy_latents, args.min_snr_gamma)
 
                 loss = loss.mean()  # 平均なのでbatch_sizeで割る必要なし
 
@@ -662,6 +656,7 @@ def setup_parser() -> argparse.ArgumentParser:
     train_util.add_training_arguments(parser, True)
     train_util.add_optimizer_arguments(parser)
     config_util.add_config_arguments(parser)
+    custom_train_functions.add_custom_train_arguments(parser)
 
     parser.add_argument("--no_metadata", action="store_true", help="do not save metadata in output model / メタデータを出力先モデルに保存しない")
     parser.add_argument(

train_textual_inversion.py

@@ -17,6 +17,8 @@
     ConfigSanitizer,
     BlueprintGenerator,
 )
+import library.custom_train_functions as custom_train_functions
+from library.custom_train_functions import apply_snr_weight
 
 imagenet_templates_small = [
     "a photo of a {}",
@@ -377,6 +379,9 @@ def train(args):
 
                 loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float(), reduction="none")
                 loss = loss.mean([1, 2, 3])
+
+                if args.min_snr_gamma:
+                  loss = apply_snr_weight(loss, latents, noisy_latents, args.min_snr_gamma)
 
                 loss_weights = batch["loss_weights"]  # 各sampleごとのweight
                 loss = loss * loss_weights
@@ -534,6 +539,7 @@ def setup_parser() -> argparse.ArgumentParser:
     train_util.add_training_arguments(parser, True)
     train_util.add_optimizer_arguments(parser)
     config_util.add_config_arguments(parser)
+    custom_train_functions.add_custom_train_arguments(parser)
 
     parser.add_argument(
         "--save_model_as",