Min-SNR Weighting Strategy: Refactored and added to all trainers

train_db.py

@@ -21,7 +21,8 @@ from library.config_util import (
     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",