Update

5 files changed, 64 insertions, 10 deletions

diff --git a/train_dreambooth.py b/train_dreambooth.py
index 929310b..90ca467 100644
--- a/train_dreambooth.py
+++ b/train_dreambooth.py
@@ -839,7 +839,10 @@ def main():
 
         create_optimizer = partial(
             prodigyopt.Prodigy,
+            betas=(args.adam_beta1, args.adam_beta2),
             weight_decay=args.adam_weight_decay,
+            eps=args.adam_epsilon,
+            d0=args.dadaptation_d0,
         )
 
         args.learning_rate_unet = 1.0
@@ -965,9 +968,23 @@ def main():
         },
         {
             "params": (
-                param for param in text_encoder.parameters() if param.requires_grad
+                param
+                for param in itertools.chain(
+                    text_encoder.text_model.encoder.parameters(),
+                    text_encoder.text_model.final_layer_norm.parameters(),
+                )
+                if param.requires_grad
+            ),
+            "lr": learning_rate_text,
+        },
+        {
+            "params": (
+                param
+                for param in text_encoder.text_model.embeddings.token_embedding.parameters()
+                if param.requires_grad
             ),
             "lr": learning_rate_text,
+            "weight_decay": 0,
         },
     ]
     group_labels = ["unet", "text"]
diff --git a/train_ti.py b/train_ti.py
index 1d0cb6f..a7d2924 100644
--- a/train_ti.py
+++ b/train_ti.py
@@ -337,7 +337,16 @@ def parse_args():
         "--optimizer",
         type=str,
         default="adan",
-        choices=["adam", "adam8bit", "adan", "lion", "dadam", "dadan", "adafactor"],
+        choices=[
+            "adam",
+            "adam8bit",
+            "adan",
+            "lion",
+            "dadam",
+            "dadan",
+            "adafactor",
+            "prodigy",
+        ],
         help="Optimizer to use",
     )
     parser.add_argument(
@@ -819,6 +828,23 @@ def main():
             eps=args.adam_epsilon,
             d0=args.dadaptation_d0,
         )
+    elif args.optimizer == "prodigy":
+        try:
+            import prodigyopt
+        except ImportError:
+            raise ImportError(
+                "To use DAdaptAdan, please install the dadaptation library: `pip install dadaptation`."
+            )
+
+        create_optimizer = partial(
+            prodigyopt.Prodigy,
+            betas=(args.adam_beta1, args.adam_beta2),
+            weight_decay=args.adam_weight_decay,
+            eps=args.adam_epsilon,
+            d0=args.dadaptation_d0,
+        )
+
+        args.learning_rate = 1.0
     else:
         raise ValueError(f'Unknown --optimizer "{args.optimizer}"')
 
@@ -959,7 +985,11 @@ def main():
         avg_acc_val = AverageMeter()
 
         optimizer = create_optimizer(
-            text_encoder.text_model.embeddings.token_embedding.parameters(),
+            (
+                param
+                for param in text_encoder.text_model.embeddings.token_embedding.parameters()
+                if param.requires_grad
+            ),
             lr=args.learning_rate,
         )
 
@@ -973,9 +1003,11 @@ def main():
 
             if response.lower().strip() == "o":
                 if args.learning_rate is not None:
-                    learning_rate = args.learning_rate * 2
+                    learning_rate = (
+                        args.learning_rate * 2 * (args.cycle_decay**training_iter)
+                    )
             else:
-                learning_rate = args.learning_rate
+                learning_rate = args.learning_rate * (args.cycle_decay**training_iter)
 
             if response.lower().strip() == "o":
                 lr_scheduler = "one_cycle"
@@ -1045,8 +1077,6 @@ def main():
             )
 
             training_iter += 1
-            if learning_rate is not None:
-                learning_rate *= args.cycle_decay
 
         accelerator.end_training()
 
diff --git a/training/functional.py b/training/functional.py
index 8917eb7..b60afe3 100644
--- a/training/functional.py
+++ b/training/functional.py
@@ -786,7 +786,4 @@ def train(
     accelerator.unwrap_model(text_encoder, keep_fp32_wrapper=False)
     accelerator.unwrap_model(unet, keep_fp32_wrapper=False)
 
-    text_encoder.forward = MethodType(text_encoder.forward, text_encoder)
-    unet.forward = MethodType(unet.forward, unet)
-
     accelerator.free_memory()
diff --git a/training/strategy/dreambooth.py b/training/strategy/dreambooth.py
index 3d1abf7..7e67589 100644
--- a/training/strategy/dreambooth.py
+++ b/training/strategy/dreambooth.py
@@ -154,6 +154,9 @@ def dreambooth_strategy_callbacks(
         unet_ = accelerator.unwrap_model(unet, keep_fp32_wrapper=False)
         text_encoder_ = accelerator.unwrap_model(text_encoder, keep_fp32_wrapper=False)
 
+        text_encoder_.forward = MethodType(text_encoder_.forward, text_encoder_)
+        unet_.forward = MethodType(unet_.forward, unet_)
+
         text_encoder_.text_model.embeddings.persist(False)
 
         with ema_context():
diff --git a/training/strategy/ti.py b/training/strategy/ti.py
index 7373982..f37dfb4 100644
--- a/training/strategy/ti.py
+++ b/training/strategy/ti.py
@@ -1,4 +1,5 @@
 from typing import Optional
+from types import MethodType
 from functools import partial
 from contextlib import contextmanager, nullcontext
 from pathlib import Path
@@ -139,6 +140,12 @@ def textual_inversion_strategy_callbacks(
     def on_checkpoint(step, postfix):
         print(f"Saving checkpoint for step {step}...")
 
+        if postfix == "end":
+            text_encoder_ = accelerator.unwrap_model(
+                text_encoder, keep_fp32_wrapper=False
+            )
+            text_encoder_.forward = MethodType(text_encoder_.forward, text_encoder_)
+
         with ema_context():
             for token, ids in zip(placeholder_tokens, placeholder_token_ids):
                 text_encoder.text_model.embeddings.save_embed(