4 files changed, 34 insertions, 14 deletions

diff --git a/training/functional.py b/training/functional.py
index 2da0f69..ebc40de 100644
--- a/training/functional.py
+++ b/training/functional.py
@@ -42,7 +42,7 @@ class TrainingCallbacks():
     on_after_optimize: Callable[[Any, dict[str, float]], None] = const()
     on_after_epoch: Callable[[], None] = const()
     on_eval: Callable[[], _GeneratorContextManager] = const(nullcontext())
-    on_sample: Callable[[int], None] = const()
+    on_sample: Callable[[int, int], None] = const()
     on_checkpoint: Callable[[int, str], None] = const()
 
 
@@ -96,6 +96,7 @@ def save_samples(
     output_dir: Path,
     seed: int,
     step: int,
+    cycle: int = 1,
     batch_size: int = 1,
     num_batches: int = 1,
     num_steps: int = 20,
@@ -125,7 +126,7 @@ def save_samples(
 
     for pool, data, gen in datasets:
         all_samples = []
-        file_path = output_dir / pool / f"step_{step}.jpg"
+        file_path = output_dir / pool / f"step_{cycle}_{step}.jpg"
         file_path.parent.mkdir(parents=True, exist_ok=True)
 
         batches = list(itertools.islice(itertools.cycle(data), batch_size * num_batches))
@@ -455,7 +456,7 @@ def train_loop(
     sample_frequency: int = 10,
     checkpoint_frequency: int = 50,
     milestone_checkpoints: bool = True,
-    initial_samples: bool = True,
+    cycle: int = 1,
     global_step_offset: int = 0,
     num_epochs: int = 100,
     gradient_accumulation_steps: int = 1,
@@ -518,12 +519,12 @@ def train_loop(
     try:
         for epoch in range(num_epochs):
             if accelerator.is_main_process:
-                if epoch % sample_frequency == 0 and (initial_samples or epoch != 0):
+                if epoch % sample_frequency == 0 and (cycle == 1 or epoch != 0):
                     local_progress_bar.clear()
                     global_progress_bar.clear()
 
                     with on_eval():
-                        on_sample(global_step)
+                        on_sample(cycle, global_step)
 
                 if epoch % checkpoint_frequency == 0 and epoch != 0:
                     local_progress_bar.clear()
@@ -648,7 +649,7 @@ def train_loop(
         if accelerator.is_main_process:
             print("Finished!")
             with on_eval():
-                on_sample(global_step)
+                on_sample(cycle, global_step)
             on_checkpoint(global_step, "end")
 
     except KeyboardInterrupt:
@@ -680,7 +681,7 @@ def train(
     sample_frequency: int = 20,
     checkpoint_frequency: int = 50,
     milestone_checkpoints: bool = True,
-    initial_samples: bool = True,
+    cycle: int = 1,
     global_step_offset: int = 0,
     guidance_scale: float = 0.0,
     prior_loss_weight: float = 1.0,
@@ -731,7 +732,7 @@ def train(
         sample_frequency=sample_frequency,
         checkpoint_frequency=checkpoint_frequency,
         milestone_checkpoints=milestone_checkpoints,
-        initial_samples=initial_samples,
+        cycle=cycle,
         global_step_offset=global_step_offset,
         num_epochs=num_train_epochs,
         gradient_accumulation_steps=gradient_accumulation_steps,
diff --git a/training/strategy/dreambooth.py b/training/strategy/dreambooth.py
index 4ae28b7..e6fcc89 100644
--- a/training/strategy/dreambooth.py
+++ b/training/strategy/dreambooth.py
@@ -148,7 +148,7 @@ def dreambooth_strategy_callbacks(
             torch.cuda.empty_cache()
 
     @torch.no_grad()
-    def on_sample(step):
+    def on_sample(cycle, step):
         unet_ = accelerator.unwrap_model(unet, keep_fp32_wrapper=True)
         text_encoder_ = accelerator.unwrap_model(text_encoder, keep_fp32_wrapper=True)
 
@@ -158,7 +158,7 @@ def dreambooth_strategy_callbacks(
         unet_.to(dtype=weight_dtype)
         text_encoder_.to(dtype=weight_dtype)
 
-        save_samples_(step=step, unet=unet_, text_encoder=text_encoder_)
+        save_samples_(cycle=cycle, step=step, unet=unet_, text_encoder=text_encoder_)
 
         unet_.to(dtype=orig_unet_dtype)
         text_encoder_.to(dtype=orig_text_encoder_dtype)
diff --git a/training/strategy/lora.py b/training/strategy/lora.py
index 48236fb..5c3012e 100644
--- a/training/strategy/lora.py
+++ b/training/strategy/lora.py
@@ -146,11 +146,11 @@ def lora_strategy_callbacks(
             torch.cuda.empty_cache()
 
     @torch.no_grad()
-    def on_sample(step):
+    def on_sample(cycle, step):
         unet_ = accelerator.unwrap_model(unet, keep_fp32_wrapper=True)
         text_encoder_ = accelerator.unwrap_model(text_encoder, keep_fp32_wrapper=True)
 
-        save_samples_(step=step, unet=unet_, text_encoder=text_encoder_)
+        save_samples_(cycle=cycle, step=step, unet=unet_, text_encoder=text_encoder_)
 
         del unet_, text_encoder_
 
diff --git a/training/strategy/ti.py b/training/strategy/ti.py
index f0b84b5..6bbff64 100644
--- a/training/strategy/ti.py
+++ b/training/strategy/ti.py
@@ -104,10 +104,28 @@ def textual_inversion_strategy_callbacks(
             yield
 
     @torch.no_grad()
+    def on_before_optimize(epoch: int):
+        if use_emb_decay:
+            params = [
+                p
+                for p in text_encoder.text_model.embeddings.token_embedding.parameters()
+                if p.grad is not None
+            ]
+            return torch.stack(params) if len(params) != 0 else None
+
+    @torch.no_grad()
     def on_after_optimize(w, lrs: dict[str, float]):
         if ema_embeddings is not None:
             ema_embeddings.step(text_encoder.text_model.embeddings.token_embedding.parameters())
 
+        if use_emb_decay and w is not None:
+            lr = lrs["emb"] or lrs["0"]
+            lambda_ = emb_decay * lr
+
+            if lambda_ != 0:
+                norm = w[:, :].norm(dim=-1, keepdim=True)
+                w[:].add_((w[:] / norm.clamp_min(1e-12)) * lambda_ * (emb_decay_target - norm))
+
     def on_log():
         if ema_embeddings is not None:
             return {"ema_decay": ema_embeddings.decay}
@@ -125,7 +143,7 @@ def textual_inversion_strategy_callbacks(
                 )
 
     @torch.no_grad()
-    def on_sample(step):
+    def on_sample(cycle, step):
         unet_ = accelerator.unwrap_model(unet, keep_fp32_wrapper=True)
         text_encoder_ = accelerator.unwrap_model(text_encoder, keep_fp32_wrapper=True)
 
@@ -135,7 +153,7 @@ def textual_inversion_strategy_callbacks(
         unet_.to(dtype=weight_dtype)
         text_encoder_.to(dtype=weight_dtype)
 
-        save_samples_(step=step, unet=unet_, text_encoder=text_encoder_)
+        save_samples_(cycle=cycle, step=step, unet=unet_, text_encoder=text_encoder_)
 
         unet_.to(dtype=orig_unet_dtype)
         text_encoder_.to(dtype=orig_text_encoder_dtype)
@@ -148,6 +166,7 @@ def textual_inversion_strategy_callbacks(
     return TrainingCallbacks(
         on_train=on_train,
         on_eval=on_eval,
+        on_before_optimize=on_before_optimize,
         on_after_optimize=on_after_optimize,
         on_log=on_log,
         on_checkpoint=on_checkpoint,