Merge branch 'dev' into master

dreambooth_gui.py

@@ -25,7 +25,7 @@
     gradio_config,
     gradio_source_model,
     set_legacy_8bitadam,
-    update_optimizer,
+    my_data,
 )
 from library.tensorboard_gui import (
     gradio_tensorboard,
@@ -214,7 +214,7 @@ def open_configuration(
             my_data = json.load(f)
             print('Loading config...')
             # Update values to fix deprecated use_8bit_adam checkbox and set appropriate optimizer if it is set to True
-            my_data = update_optimizer(my_data)
+            my_data = my_data(my_data)
     else:
         file_path = original_file_path  # In case a file_path was provided and the user decide to cancel the open action
         my_data = {}

finetune_gui.py

@@ -19,7 +19,7 @@
     color_aug_changed,
     run_cmd_training,
     set_legacy_8bitadam,
-    update_optimizer,
+    my_data,
 )
 from library.tensorboard_gui import (
     gradio_tensorboard,
@@ -217,7 +217,7 @@ def open_config_file(
             my_data_db = json.load(f)
             print('Loading config...')
             # Update values to fix deprecated use_8bit_adam checkbox and set appropriate optimizer if it is set to True
-            my_data = update_optimizer(my_data)
+            my_data = my_data(my_data)
     else:
         file_path = original_file_path  # In case a file_path was provided and the user decide to cancel the open action
         my_data_db = {}

library/common_gui.py

@@ -10,10 +10,14 @@
 document_symbol = '\U0001F4C4'   # 📄
 
 
-def update_optimizer(my_data):
+def my_data(my_data):
     if my_data.get('use_8bit_adam', False):
         my_data['optimizer'] = 'AdamW8bit'
         my_data['use_8bit_adam'] = False
+
+    if my_data.get('model_list', 'custom') == []:
+        print('Old config with empty model list. Setting to custom...')
+        my_data['model_list'] = 'custom'
     return my_data
 
 

library/git_caption_gui.py

@@ -25,7 +25,7 @@ def caption_images(
         return
 
     print(f'GIT captioning files in {train_data_dir}...')
-    run_cmd = f'{PYTHON} "finetune/make_captions.py"'
+    run_cmd = f'{PYTHON} "finetune/make_captions_by_git.py"'
     if not model_id == '':
         run_cmd += f' --model_id="{model_id}"'
     run_cmd += f' --batch_size="{int(batch_size)}"'

lora_gui.py

@@ -25,7 +25,7 @@
     gradio_source_model,
     run_cmd_training,
     set_legacy_8bitadam,
-    update_optimizer,
+    my_data,
 )
 from library.dreambooth_folder_creation_gui import (
     gradio_dreambooth_folder_creation_tab,
@@ -239,7 +239,7 @@ def open_configuration(
             my_data = json.load(f)
             print('Loading config...')
             # Update values to fix deprecated use_8bit_adam checkbox and set appropriate optimizer if it is set to True
-            my_data = update_optimizer(my_data)
+            my_data = my_data(my_data)
     else:
         file_path = original_file_path  # In case a file_path was provided and the user decide to cancel the open action
         my_data = {}

textual_inversion_gui.py

@@ -25,7 +25,7 @@
     gradio_config,
     gradio_source_model,
     set_legacy_8bitadam,
-    update_optimizer,
+    my_data,
 )
 from library.tensorboard_gui import (
     gradio_tensorboard,
@@ -226,7 +226,7 @@ def open_configuration(
             my_data_db = json.load(f)
             print('Loading config...')
             # Update values to fix deprecated use_8bit_adam checkbox and set appropriate optimizer if it is set to True
-            my_data = update_optimizer(my_data)
+            my_data = my_data(my_data)
     else:
         file_path = original_file_path  # In case a file_path was provided and the user decide to cancel the open action
         my_data_db = {}

tools/extract_locon.py

@@ -0,0 +1,106 @@
+#
+# From: https://raw.githubusercontent.com/KohakuBlueleaf/LoCon/main/extract_locon.py
+#
+
+import argparse
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "base_model", help="The model which use it to train the dreambooth model",
+        default='', type=str
+    )
+    parser.add_argument(
+        "db_model", help="the dreambooth model you want to extract the locon",
+        default='', type=str
+    )
+    parser.add_argument(
+        "output_name", help="the output model",
+        default='./out.pt', type=str
+    )
+    parser.add_argument(
+        "--is_v2", help="Your base/db model is sd v2 or not",
+        default=False, action="store_true"
+    )
+    parser.add_argument(
+        "--device", help="Which device you want to use to extract the locon",
+        default='cpu', type=str
+    )
+    parser.add_argument(
+        "--mode", 
+        help=(
+            'extraction mode, can be "fixed", "threshold", "ratio", "percentile". '
+            'If not "fixed", network_dim and conv_dim will be ignored'
+        ),
+        default='fixed', type=str
+    )
+    parser.add_argument(
+        "--linear_dim", help="network dim for linear layer in fixed mode",
+        default=1, type=int
+    )
+    parser.add_argument(
+        "--conv_dim", help="network dim for conv layer in fixed mode",
+        default=1, type=int
+    )
+    parser.add_argument(
+        "--linear_threshold", help="singular value threshold for linear layer in threshold mode",
+        default=0., type=float
+    )
+    parser.add_argument(
+        "--conv_threshold", help="singular value threshold for conv layer in threshold mode",
+        default=0., type=float
+    )
+    parser.add_argument(
+        "--linear_ratio", help="singular ratio for linear layer in ratio mode",
+        default=0., type=float
+    )
+    parser.add_argument(
+        "--conv_ratio", help="singular ratio for conv layer in ratio mode",
+        default=0., type=float
+    )
+    parser.add_argument(
+        "--linear_percentile", help="singular value percentile for linear layer percentile mode",
+        default=1., type=float
+    )
+    parser.add_argument(
+        "--conv_percentile", help="singular value percentile for conv layer percentile mode",
+        default=1., type=float
+    )
+    return parser.parse_args()
+ARGS = get_args()
+
+from locon.utils import extract_diff
+from locon.kohya_model_utils import load_models_from_stable_diffusion_checkpoint
+
+import torch
+
+
+def main():
+    args = ARGS
+    base = load_models_from_stable_diffusion_checkpoint(args.is_v2, args.base_model)
+    db = load_models_from_stable_diffusion_checkpoint(args.is_v2, args.db_model)
+
+    linear_mode_param = {
+        'fixed': args.linear_dim,
+        'threshold': args.linear_threshold,
+        'ratio': args.linear_ratio,
+        'percentile': args.linear_percentile,
+    }[args.mode]
+    conv_mode_param = {
+        'fixed': args.conv_dim,
+        'threshold': args.conv_threshold,
+        'ratio': args.conv_ratio,
+        'percentile': args.conv_percentile,
+    }[args.mode]
+
+    state_dict = extract_diff(
+        base, db,
+        args.mode,
+        linear_mode_param, conv_mode_param,
+        args.device
+    )
+    torch.save(state_dict, args.output_name)
+
+
+if __name__ == '__main__':
+    main()