Fix/leco cleanup (#2294)

* feat: SD1.x/2.x と SDXL 向けの LECO 学習スクリプトを追加 (#2285)

* Add LECO training script and associated tests

- Implemented `sdxl_train_leco.py` for training with LECO prompts, including argument parsing, model setup, training loop, and weight saving functionality.
- Created unit tests for `load_prompt_settings` in `test_leco_train_util.py` to validate loading of prompt configurations in both original and slider formats.
- Added basic syntax tests for `train_leco.py` and `sdxl_train_leco.py` to ensure modules are importable.

* fix: use getattr for safe attribute access in argument verification

* feat: add CUDA device compatibility validation and corresponding tests

* Revert "feat: add CUDA device compatibility validation and corresponding tests"

This reverts commit 6d3e51431b.

* feat: update predict_noise_xl to use vector embedding from add_time_ids

* feat: implement checkpointing in predict_noise and predict_noise_xl functions

* feat: remove unused submodules and update .gitignore to exclude .codex-tmp

---------

Co-authored-by: Kohya S. <52813779+kohya-ss@users.noreply.github.com>

* fix: format

* fix: LECO PR #2285 のレビュー指摘事項を修正

- train_util.py/deepspeed_utils.py の getattr 化を元に戻し、LECO パーサーにダミー引数を追加
- sdxl_train_util のモジュールレベルインポートをローカルインポートに変更
- PromptEmbedsCache.__getitem__ でキャッシュミス時に KeyError を送出するよう修正
- 設定ファイル形式を YAML から TOML に変更(リポジトリの規約に統一)
- 重複コード (build_network_kwargs, get_save_extension, save_weights) を leco_train_util.py に統合
- _expand_slider_target の冗長な PromptSettings 構築を簡素化
- add_time_ids 用に専用の batch_add_time_ids 関数を追加

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* docs: LECO 学習ガイドを大幅に拡充

コマンドライン引数の全カテゴリ別解説、プロンプト TOML の全フィールド説明、
2つの guidance_scale の違い、推奨設定表、YAML からの変換ガイド等を追加。
英語本文と日本語折り畳みの二言語構成。

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: apply_noise_offset の dtype 不一致を修正

torch.randn のデフォルト float32 により latents が暗黙的にアップキャストされる問題を修正。
float32/CPU で生成後に latents の dtype/device へ変換する安全なパターンを採用。

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Umisetokikaze <52318966+umisetokikaze@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
Kohya S.
2026-03-29 20:41:43 +09:00
committed by GitHub
parent 89b246f3f6
commit 5cdad10de5
9 changed files with 2074 additions and 6 deletions

View File

@@ -1106,7 +1106,8 @@ class BaseDataset(torch.utils.data.Dataset):
return all(
[
not (
subset.caption_dropout_rate > 0 and not cache_supports_dropout
subset.caption_dropout_rate > 0
and not cache_supports_dropout
or subset.shuffle_caption
or subset.token_warmup_step > 0
or subset.caption_tag_dropout_rate > 0
@@ -2056,7 +2057,9 @@ class DreamBoothDataset(BaseDataset):
filtered_img_paths.append(img_path)
filtered_sizes.append(size)
if len(filtered_img_paths) < len(img_paths):
logger.info(f"filtered {len(img_paths) - len(filtered_img_paths)} images by original resolution from {subset.image_dir}")
logger.info(
f"filtered {len(img_paths) - len(filtered_img_paths)} images by original resolution from {subset.image_dir}"
)
img_paths = filtered_img_paths
sizes = filtered_sizes
@@ -2542,9 +2545,7 @@ class ControlNetDataset(BaseDataset):
len(missing_imgs) == 0
), f"missing conditioning data for {len(missing_imgs)} images / 制御用画像が見つかりませんでした: {missing_imgs}"
if len(extra_imgs) > 0:
logger.warning(
f"extra conditioning data for {len(extra_imgs)} images / 余分な制御用画像があります: {extra_imgs}"
)
logger.warning(f"extra conditioning data for {len(extra_imgs)} images / 余分な制御用画像があります: {extra_imgs}")
self.conditioning_image_transforms = IMAGE_TRANSFORMS