Remove the skew and kurtosis loss objectives; they hurt more than the…

…y help
cheald · Apr 24, 2024 · 357aa44 · 357aa44
1 parent 6bf5dae
commit 357aa44
Show file tree

Hide file tree

Showing 2 changed files with 0 additions and 27 deletions.
diff --git a/library/train_util.py b/library/train_util.py
@@ -3369,25 +3369,6 @@ def add_training_arguments(parser: argparse.ArgumentParser, support_dreambooth:
         default=None,
         help="Weight for standard deviation loss. Encourages the model to learn noise with a stddev like the true noise. May prevent 'deep fry'. 1.0 is a good starting place.",
     )
-    parser.add_argument(
-        "--kurtosis_loss_weight",
-        type=float,
-        default=None,
-        help="Weight for kurtosis loss. Encourages the model to learn noise with a kurtosis like the true noise. Recommended if using std_loss_weight.",
-    )
-    parser.add_argument(
-        "--skew_loss_weight",
-        type=float,
-        default=None,
-        help="Weight for skew loss. Encourages the model to learn noise with a skew like the true noise. Recommended if using std_loss_weight.",
-    )
-    parser.add_argument(
-        "--latent_corruption",
-        type=float,
-        default=None,
-        help="latent corruption for training (default is None) / 学習時のlatent corruption（デフォルトはNone）",
-    )
-
 
     if support_dreambooth:
         # DreamBooth training

diff --git a/train_network.py b/train_network.py
@@ -913,14 +913,6 @@ def remove_model(old_ckpt_name):
                         std_loss = F.mse_loss(pred_std, true_std, reduction="none")
                         loss = loss + std_loss * args.std_loss_weight
 
-                    if args.skew_loss_weight is not None:
-                        skew_loss = F.mse_loss(pred_skews, true_skews, reduction="none")
-                        loss = loss + skew_loss * args.skew_loss_weight
-
-                    if args.kurtosis_loss_weight is not None:
-                        kurtosis_loss = F.mse_loss(pred_kurtoses, true_kurtoses, reduction="none")
-                        loss = loss + kurtosis_loss * args.kurtosis_loss_weight
-
                     # print(kl_loss.dtype, pred_std.dtype, noise_pred.dtype, true_std.dtype, pred_skews.dtype, true_skews.dtype, pred_kurtoses.dtype, true_kurtoses.dtype)
                     # step_logs["loss/kl_loss"]                = kl_loss.mean().item()
                     step_logs["metrics/noise_pred_std"]      = pred_std.mean().item()