small tweaks

preprocess.py

@@ -196,7 +196,7 @@ def clipseg_mask_generator(
 
     if isinstance(target_prompts, str):
         print(
-            f'Warning: only one target prompt "{target_prompts}" was given, so it will be used for all images'
+            f'Using "{target_prompts}" as CLIP-segmentation prompt for all images.'
         )
         target_prompts = [target_prompts] * len(images)
 

trainer/utils/config_modification.py

@@ -1,28 +1,21 @@
 from ..config import TrainingConfig
 
-def modify_args_based_on_concept_mode(
-    concept_mode: str,
-    left_right_flip_augmentation: bool,
-    mask_target_prompts: str,
-    clipseg_temperature: float,
-    l1_penalty: float
-):
-    if concept_mode == "face":
+def modify_args_based_on_concept_mode(config: TrainingConfig):
+    if config.concept_mode == "face":
         print(f"Face mode is active ----> disabling left-right flips and setting mask_target_prompts to 'face'.")
-        left_right_flip_augmentation = False  # always disable lr flips for face mode!
-        mask_target_prompts = "face"
-        clipseg_temperature = 0.4
+        config.left_right_flip_augmentation = False  # always disable lr flips for face mode!
+        config.mask_target_prompts = "face"
+        config.clipseg_temperature = 0.4
 
-    if concept_mode == "concept": # gracefully catch any old versions of concept_mode
-        concept_mode = "object"
+    if config.concept_mode == "concept": # gracefully catch any old versions of concept_mode
+        config.concept_mode = "object"
 
-    if concept_mode == "style": # for styles you usually want the LoRA matrices to absorb a lot (instead of just the token embedding)
-        l1_penalty = 0.05
+    if config.concept_mode == "style": # for styles you usually want the LoRA matrices to absorb a lot (instead of just the token embedding)
+        config.l1_penalty = 0.05
 
-    return (
-        concept_mode,
-        left_right_flip_augmentation,
-        mask_target_prompts,
-        clipseg_temperature,
-        l1_penalty
-    )
+    if config.use_dora:
+        print(f"Disabling L1 penalty and LoRA weight decay for DORA training.")
+        config.l1_penalty = 0.0
+        config.lora_weight_decay = 0.0
+
+    return config

trainer_pti.py

@@ -70,14 +70,7 @@ def main(
     seed_everything(config.seed)
     pick_best_gpu_id()
 
-    (config.concept_mode, config.left_right_flip_augmentation, config.mask_target_prompts, config.clipseg_temperature, config.l1_penalty
-        ) = modify_args_based_on_concept_mode(
-            concept_mode=config.concept_mode,
-            left_right_flip_augmentation=config.left_right_flip_augmentation,
-            mask_target_prompts=config.mask_target_prompts,
-            clipseg_temperature=config.clipseg_temperature,
-            l1_penalty=config.l1_penalty
-        )
+    config = modify_args_based_on_concept_mode(config)
 
     input_dir, n_imgs, trigger_text, segmentation_prompt, captions = preprocess(
         working_directory=config.output_dir,

training_args.json

@@ -5,7 +5,7 @@
     "lora_training_urls": "https://storage.googleapis.com/public-assets-xander/A_workbox/lora_training_sets/clipx_tiny.zip",
     "concept_mode": "style",
     "caption_prefix": "in the style of TOK, ",
-    "seed": 0,
+    "seed": 1,
     "resolution": 1024,
     "train_batch_size": 4,
     "max_train_steps": 500,