Update

3 files changed, 55 insertions, 9 deletions

diff --git a/train_dreambooth.py b/train_dreambooth.py
index 05f6cb5..1e49474 100644
--- a/train_dreambooth.py
+++ b/train_dreambooth.py
@@ -565,8 +565,6 @@ def main():
     checkpoint_scheduler = DPMSolverMultistepScheduler.from_pretrained(
         args.pretrained_model_name_or_path, subfolder='scheduler')
 
-    tokenizer.set_use_vector_shuffle(args.vector_shuffle)
-
     vae.enable_slicing()
     vae.set_use_memory_efficient_attention_xformers(True)
     unet.set_use_memory_efficient_attention_xformers(True)
@@ -893,7 +891,16 @@ def main():
         accelerator.init_trackers("dreambooth", config=config)
 
     if args.find_lr:
-        lr_finder = LRFinder(accelerator, text_encoder, optimizer, train_dataloader, val_dataloader, loop)
+        lr_finder = LRFinder(
+            accelerator,
+            text_encoder,
+            optimizer,
+            train_dataloader,
+            val_dataloader,
+            loop,
+            on_train=lambda: tokenizer.set_use_vector_shuffle(args.vector_shuffle),
+            on_eval=lambda: tokenizer.set_use_vector_shuffle(False)
+        )
         lr_finder.run(min_lr=1e-4)
 
         plt.savefig(basepath.joinpath("lr.png"))
@@ -965,11 +972,11 @@ def main():
             local_progress_bar.reset()
 
             unet.train()
-
             if epoch < args.train_text_encoder_epochs:
                 text_encoder.train()
             elif epoch == args.train_text_encoder_epochs:
                 text_encoder.requires_grad_(False)
+            tokenizer.set_use_vector_shuffle(args.vector_shuffle)
 
             for step, batch in enumerate(train_dataloader):
                 with accelerator.accumulate(unet):
@@ -1023,6 +1030,7 @@ def main():
 
             unet.eval()
             text_encoder.eval()
+            tokenizer.set_use_vector_shuffle(False)
 
             cur_loss_val = AverageMeter()
             cur_acc_val = AverageMeter()
diff --git a/train_ti.py b/train_ti.py
index 97dde1e..2b3f017 100644
--- a/train_ti.py
+++ b/train_ti.py
@@ -535,8 +535,6 @@ def main():
     checkpoint_scheduler = DPMSolverMultistepScheduler.from_pretrained(
         args.pretrained_model_name_or_path, subfolder='scheduler')
 
-    tokenizer.set_use_vector_shuffle(args.vector_shuffle)
-
     vae.enable_slicing()
     vae.set_use_memory_efficient_attention_xformers(True)
     unet.set_use_memory_efficient_attention_xformers(True)
@@ -845,7 +843,16 @@ def main():
         accelerator.init_trackers("textual_inversion", config=config)
 
     if args.find_lr:
-        lr_finder = LRFinder(accelerator, text_encoder, optimizer, train_dataloader, val_dataloader, loop)
+        lr_finder = LRFinder(
+            accelerator,
+            text_encoder,
+            optimizer,
+            train_dataloader,
+            val_dataloader,
+            loop,
+            on_train=lambda: tokenizer.set_use_vector_shuffle(args.vector_shuffle),
+            on_eval=lambda: tokenizer.set_use_vector_shuffle(False)
+        )
         lr_finder.run(min_lr=1e-4)
 
         plt.savefig(basepath.joinpath("lr.png"))
@@ -915,6 +922,7 @@ def main():
             local_progress_bar.reset()
 
             text_encoder.train()
+            tokenizer.set_use_vector_shuffle(args.vector_shuffle)
 
             for step, batch in enumerate(train_dataloader):
                 with accelerator.accumulate(text_encoder):
@@ -955,6 +963,7 @@ def main():
             accelerator.wait_for_everyone()
 
             text_encoder.eval()
+            tokenizer.set_use_vector_shuffle(False)
 
             cur_loss_val = AverageMeter()
             cur_acc_val = AverageMeter()
diff --git a/training/lr.py b/training/lr.py
index 3abd2f2..fe166ed 100644
--- a/training/lr.py
+++ b/training/lr.py
@@ -1,5 +1,6 @@
 import math
 import copy
+from typing import Callable
 
 import matplotlib.pyplot as plt
 import numpy as np
@@ -10,19 +11,45 @@ from tqdm.auto import tqdm
 from training.util import AverageMeter
 
 
+def noop():
+    pass
+
+
 class LRFinder():
-    def __init__(self, accelerator, model, optimizer, train_dataloader, val_dataloader, loss_fn):
+    def __init__(
+        self,
+        accelerator,
+        model,
+        optimizer,
+        train_dataloader,
+        val_dataloader,
+        loss_fn,
+        on_train: Callable[[], None] = noop,
+        on_eval: Callable[[], None] = noop
+    ):
         self.accelerator = accelerator
         self.model = model
         self.optimizer = optimizer
         self.train_dataloader = train_dataloader
         self.val_dataloader = val_dataloader
         self.loss_fn = loss_fn
+        self.on_train = on_train
+        self.on_eval = on_eval
 
         # self.model_state = copy.deepcopy(model.state_dict())
         # self.optimizer_state = copy.deepcopy(optimizer.state_dict())
 
-    def run(self, min_lr, skip_start=10, skip_end=5, num_epochs=100, num_train_batches=1, num_val_batches=math.inf, smooth_f=0.05, diverge_th=5):
+    def run(
+        self,
+        min_lr,
+        skip_start: int = 10,
+        skip_end: int = 5,
+        num_epochs: int = 100,
+        num_train_batches: int = 1,
+        num_val_batches: int = math.inf,
+        smooth_f: float = 0.05,
+        diverge_th: int = 5
+    ):
         best_loss = None
         best_acc = None
 
@@ -50,6 +77,7 @@ class LRFinder():
             avg_acc = AverageMeter()
 
             self.model.train()
+            self.on_train()
 
             for step, batch in enumerate(self.train_dataloader):
                 if step >= num_train_batches:
@@ -67,6 +95,7 @@ class LRFinder():
                     progress_bar.update(1)
 
             self.model.eval()
+            self.on_eval()
 
             with torch.inference_mode():
                 for step, batch in enumerate(self.val_dataloader):