Update

4 files changed, 12 insertions, 8 deletions

diff --git a/train_lora.py b/train_lora.py
index 0d8b8cb..1d1485d 100644
--- a/train_lora.py
+++ b/train_lora.py
@@ -873,7 +873,6 @@ def main():
         seed=args.seed,
         guidance_scale=args.guidance_scale,
         prior_loss_weight=args.prior_loss_weight if args.num_class_images != 0 else 0,
-        offset_noise_strength=args.offset_noise_strength,
         sample_scheduler=sample_scheduler,
         sample_batch_size=args.sample_batch_size,
         sample_num_batches=args.sample_batches,
@@ -984,13 +983,14 @@ def main():
             lr_scheduler=pti_lr_scheduler,
             num_train_epochs=num_train_epochs,
             gradient_accumulation_steps=args.gradient_accumulation_steps,
-            cycle=1,
+            cycle=0,
             pti_mode=True,
             # --
             group_labels=["emb"],
             sample_output_dir=pti_sample_output_dir,
             checkpoint_output_dir=pti_checkpoint_output_dir,
             sample_frequency=pti_sample_frequency,
+            offset_noise_strength=0,
             no_val=True,
         )
 
@@ -1132,11 +1132,13 @@ def main():
             gradient_accumulation_steps=args.gradient_accumulation_steps,
             global_step_offset=training_iter * num_train_steps,
             cycle=training_iter,
+            train_text_encoder_cycles=args.train_text_encoder_cycles,
             # --
             group_labels=group_labels,
             sample_output_dir=lora_sample_output_dir,
             checkpoint_output_dir=lora_checkpoint_output_dir,
             sample_frequency=lora_sample_frequency,
+            offset_noise_strength=args.offset_noise_strength,
             no_val=args.valid_set_size == 0,
         )
 
diff --git a/training/functional.py b/training/functional.py
index c6ceb20..695a24f 100644
--- a/training/functional.py
+++ b/training/functional.py
@@ -456,7 +456,7 @@ def train_loop(
     sample_frequency: int = 10,
     checkpoint_frequency: int = 50,
     milestone_checkpoints: bool = True,
-    cycle: int = 1,
+    cycle: int = 0,
     global_step_offset: int = 0,
     num_epochs: int = 100,
     gradient_accumulation_steps: int = 1,
@@ -537,7 +537,7 @@ def train_loop(
 
             logs = {}
 
-            with on_train(epoch):
+            with on_train(cycle):
                 for step, batch in enumerate(train_dataloader):
                     loss, acc, bsz = loss_step(step, batch, cache)
                     loss /= gradient_accumulation_steps
diff --git a/training/strategy/lora.py b/training/strategy/lora.py
index 5c3012e..1f0a117 100644
--- a/training/strategy/lora.py
+++ b/training/strategy/lora.py
@@ -35,6 +35,7 @@ def lora_strategy_callbacks(
     placeholder_tokens: list[str],
     placeholder_token_ids: list[list[int]],
     pti_mode: bool = False,
+    train_text_encoder_cycles: int = 99999,
     use_emb_decay: bool = False,
     emb_decay_target: float = 0.4,
     emb_decay: float = 1e-2,
@@ -66,10 +67,11 @@ def lora_strategy_callbacks(
     )
 
     @contextmanager
-    def on_train(epoch: int):
+    def on_train(cycle: int):
         unet.train()
-        text_encoder.train()
-        tokenizer.train()
+        if cycle < train_text_encoder_cycles:
+            text_encoder.train()
+            tokenizer.train()
         yield
 
     @contextmanager
diff --git a/training/strategy/ti.py b/training/strategy/ti.py
index f330cb7..6bc1d7d 100644
--- a/training/strategy/ti.py
+++ b/training/strategy/ti.py
@@ -90,7 +90,7 @@ def textual_inversion_strategy_callbacks(
             return nullcontext()
 
     @contextmanager
-    def on_train(epoch: int):
+    def on_train(cycle: int):
         text_encoder.train()
         tokenizer.train()
         yield