Merge kohya Jan 19 updates

README.md

@@ -116,6 +116,36 @@ Once you have created the LoRA network you can generate images via auto1111 by i
 
 ## Change history
 
+* 2023/01/16 (v20.3.0)
+  - Fix a part of LoRA modules are not trained when ``gradient_checkpointing`` is enabled. 
+  - Add ``--save_last_n_epochs_state`` option. You can specify how many state folders to keep, apart from how many models to keep. Thanks to shirayu!
+  - Fix Text Encoder training stops at ``max_train_steps`` even if ``max_train_epochs`` is set in `train_db.py``.
+  - Added script to check LoRA weights. You can check weights by ``python networks\check_lora_weights.py <model file>``. If some modules are not trained, the value is ``0.0`` like following. 
+    - ``lora_te_text_model_encoder_layers_11_*`` is not trained with ``clip_skip=2``, so ``0.0`` is okay for these modules.
+
+- example result of ``check_lora_weights.py``, Text Encoder and a part of U-Net are not trained:
+```
+number of LoRA-up modules: 264
+lora_te_text_model_encoder_layers_0_mlp_fc1.lora_up.weight,0.0
+lora_te_text_model_encoder_layers_0_mlp_fc2.lora_up.weight,0.0
+lora_te_text_model_encoder_layers_0_self_attn_k_proj.lora_up.weight,0.0
+:
+lora_unet_down_blocks_2_attentions_1_transformer_blocks_0_ff_net_0_proj.lora_up.weight,0.0
+lora_unet_down_blocks_2_attentions_1_transformer_blocks_0_ff_net_2.lora_up.weight,0.0
+lora_unet_mid_block_attentions_0_proj_in.lora_up.weight,0.003503334941342473
+lora_unet_mid_block_attentions_0_proj_out.lora_up.weight,0.004308608360588551
+:
+```
+
+- all modules are trained:
+```
+number of LoRA-up modules: 264
+lora_te_text_model_encoder_layers_0_mlp_fc1.lora_up.weight,0.0028684409335255623
+lora_te_text_model_encoder_layers_0_mlp_fc2.lora_up.weight,0.0029794853180646896
+lora_te_text_model_encoder_layers_0_self_attn_k_proj.lora_up.weight,0.002507600700482726
+lora_te_text_model_encoder_layers_0_self_attn_out_proj.lora_up.weight,0.002639499492943287
+:
+```
 * 2023/01/16 (v20.2.1):
     - Merging latest code update from kohya
     - Added `--max_train_epochs` and `--max_data_loader_n_workers` option for each training script.

library/train_util.py

@@ -1029,6 +1029,7 @@ def add_training_arguments(parser: argparse.ArgumentParser, support_dreambooth:
   parser.add_argument("--save_every_n_epochs", type=int, default=None,
                       help="save checkpoint every N epochs / 学習中のモデルを指定エポックごとに保存する")
   parser.add_argument("--save_last_n_epochs", type=int, default=None, help="save last N checkpoints / 最大Nエポック保存する")
+  parser.add_argument("--save_last_n_epochs_state", type=int, default=None, help="save last N checkpoints of state (overrides the value of --save_last_n_epochs)/ 最大Nエポックstateを保存する(--save_last_n_epochsの指定を上書きします)")
   parser.add_argument("--save_state", action="store_true",
                       help="save training state additionally (including optimizer states etc.) / optimizerなど学習状態も含めたstateを追加で保存する")
   parser.add_argument("--resume", type=str, default=None, help="saved state to resume training / 学習再開するモデルのstate")
@@ -1298,7 +1299,6 @@ def get_epoch_ckpt_name(args: argparse.Namespace, use_safetensors, epoch):
 
 def save_on_epoch_end(args: argparse.Namespace, save_func, remove_old_func, epoch_no: int, num_train_epochs: int):
   saving = epoch_no % args.save_every_n_epochs == 0 and epoch_no < num_train_epochs
-  remove_epoch_no = None
   if saving:
     os.makedirs(args.output_dir, exist_ok=True)
     save_func()
@@ -1306,7 +1306,7 @@ def save_on_epoch_end(args: argparse.Namespace, save_func, remove_old_func, epoc
     if args.save_last_n_epochs is not None:
       remove_epoch_no = epoch_no - args.save_every_n_epochs * args.save_last_n_epochs
       remove_old_func(remove_epoch_no)
-  return saving, remove_epoch_no
+  return saving
 
 
 def save_sd_model_on_epoch_end(args: argparse.Namespace, accelerator, src_path: str, save_stable_diffusion_format: bool, use_safetensors: bool, save_dtype: torch.dtype, epoch: int, num_train_epochs: int, global_step: int, text_encoder, unet, vae):
@@ -1346,15 +1346,18 @@ def save_sd_model_on_epoch_end(args: argparse.Namespace, accelerator, src_path:
     save_func = save_du
     remove_old_func = remove_du
 
-  saving, remove_epoch_no = save_on_epoch_end(args, save_func, remove_old_func, epoch_no, num_train_epochs)
+  saving = save_on_epoch_end(args, save_func, remove_old_func, epoch_no, num_train_epochs)
   if saving and args.save_state:
-    save_state_on_epoch_end(args, accelerator, model_name, epoch_no, remove_epoch_no)
+    save_state_on_epoch_end(args, accelerator, model_name, epoch_no)
 
 
-def save_state_on_epoch_end(args: argparse.Namespace, accelerator, model_name, epoch_no, remove_epoch_no):
+def save_state_on_epoch_end(args: argparse.Namespace, accelerator, model_name, epoch_no):
   print("saving state.")
   accelerator.save_state(os.path.join(args.output_dir, EPOCH_STATE_NAME.format(model_name, epoch_no)))
-  if remove_epoch_no is not None:
+
+  last_n_epochs = args.save_last_n_epochs_state if args.save_last_n_epochs_state else args.save_last_n_epochs
+  if last_n_epochs is not None:
+    remove_epoch_no = epoch_no - args.save_every_n_epochs * last_n_epochs
     state_dir_old = os.path.join(args.output_dir, EPOCH_STATE_NAME.format(model_name, remove_epoch_no))
     if os.path.exists(state_dir_old):
       print(f"removing old state: {state_dir_old}")

networks/check_lora_weights.py

@@ -0,0 +1,31 @@
+import argparse
+import os
+import torch
+from safetensors.torch import load_file
+
+
+def main(file):
+  print(f"loading: {file}")
+  if os.path.splitext(file)[1] == '.safetensors':
+    sd = load_file(file)
+  else:
+    sd = torch.load(file, map_location='cpu')
+
+  values = []
+
+  keys = list(sd.keys())
+  for key in keys:
+    if 'lora_up' in key:
+      values.append((key, sd[key]))
+  print(f"number of LoRA-up modules: {len(values)}")
+
+  for key, value in values:
+    print(f"{key},{torch.mean(torch.abs(value))}")
+
+
+if __name__ == '__main__':
+  parser = argparse.ArgumentParser()
+  parser.add_argument("file", type=str, help="model file to check / 重みを確認するモデルファイル")
+  args = parser.parse_args()
+
+  main(args.file)

train_db.py

@@ -92,10 +92,7 @@ def train(args):
     gc.collect()
 
   # 学習を準備する：モデルを適切な状態にする
-  if args.stop_text_encoder_training is None:
-    args.stop_text_encoder_training = args.max_train_steps + 1                # do not stop until end
-
-  train_text_encoder = args.stop_text_encoder_training >= 0
+  train_text_encoder = args.stop_text_encoder_training is None or args.stop_text_encoder_training >= 0
   unet.requires_grad_(True)                   # 念のため追加
   text_encoder.requires_grad_(train_text_encoder)
   if not train_text_encoder:
@@ -143,6 +140,9 @@ def train(args):
     args.max_train_steps = args.max_train_epochs * len(train_dataloader)
     print(f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}")
 
+  if args.stop_text_encoder_training is None:
+    args.stop_text_encoder_training = args.max_train_steps + 1                # do not stop until end
+
   # lr schedulerを用意する
   lr_scheduler = diffusers.optimization.get_scheduler(
       args.lr_scheduler, optimizer, num_warmup_steps=args.lr_warmup_steps, num_training_steps=args.max_train_steps)

train_network.py

@@ -166,6 +166,9 @@ def train(args):
   if args.gradient_checkpointing:                       # according to TI example in Diffusers, train is required
     unet.train()
     text_encoder.train()
+
+    # set top parameter requires_grad = True for gradient checkpointing works
+    text_encoder.text_model.embeddings.requires_grad_(True)
   else:
     unet.eval()
     text_encoder.eval()
@@ -364,9 +367,9 @@ def train(args):
           print(f"removing old checkpoint: {old_ckpt_file}")
           os.remove(old_ckpt_file)
 
-      saving, remove_epoch_no = train_util.save_on_epoch_end(args, save_func, remove_old_func, epoch + 1, num_train_epochs)
+      saving = train_util.save_on_epoch_end(args, save_func, remove_old_func, epoch + 1, num_train_epochs)
       if saving and args.save_state:
-        train_util.save_state_on_epoch_end(args, accelerator, model_name, epoch + 1, remove_epoch_no)
+        train_util.save_state_on_epoch_end(args, accelerator, model_name, epoch + 1)
 
     # end of epoch