Commit Graph

Select branches

Hide Pull Requests

main

Mono Color

• f9a4032783 feat: train fixed-size classifiers and export class probabilities main Shaun Zhou 2026-10-14 17:08:14 +08:00
• 63bab36147 fix: suppress normal pseudo boxes in classifier exports Shaun Zhou 2026-10-14 15:31:52 +08:00
• 219cf23abc feat: add MobileNetV3-Small classifier configuration Shaun Zhou 2026-10-14 13:29:58 +08:00
• 03f667276b feat: export classifier with YOLO OpenVINO interface Shaun Zhou 2026-10-14 12:00:36 +08:00
• d42a89ed13 fix(scripts): update default TILE_SIZE to 128 in retrain_from_label_studio.sh xzhou 2026-10-13 07:57:56 +00:00
• 8be1eefec2 fix(configs): update image size and dataset path in film.yaml to 128x128 xzhou 2026-10-13 06:53:39 +00:00
• bf72560326 docs: add AGENTS.md for vision model training and evaluation workflow Shaun Zhou 2026-10-11 16:42:13 +08:00
• 6a8941c18a docs: 修正 clusters/ 是软链的说法 Shaun ZhouandClaude Opus 5 2026-10-09 16:04:01 +08:00
• 29ff7a2c9b docs: 清掉失效的 train.py 引用与 ONNX 导出依赖组 Shaun ZhouandClaude Opus 5 2026-10-08 14:57:11 +08:00
• c2bd5301b4 feat(deploy): 模型归档统一为 film_<边长>_<时间戳> Shaun ZhouandClaude Opus 5 2026-09-30 15:11:45 +08:00
• bbd5fa167e wip: C++ demo 支持批量清单 Shaun Zhou 2026-09-29 15:07:48 +08:00
• d692ffdf6e wip: C++ 参考实现（待编译验证） Shaun Zhou 2026-09-29 14:56:21 +08:00
• ada67466d4 feat(deploy): 不依赖 ultralytics 的 OpenVINO 推理脚本 Shaun ZhouandClaude Opus 5 2026-09-24 17:50:53 +08:00
• ddd77653b5 fix(detect): 推理尺寸与训练尺寸不一致时报警 Shaun ZhouandClaude Opus 5 2026-09-24 15:33:41 +08:00
• 6e746b7351 feat(scripts): 重训后同时维护 current.pt 与 best.pt Shaun ZhouandClaude Opus 5 2026-09-23 16:03:00 +08:00
• fad1fe9fbb feat(scripts): 重训流程加入切片步骤 Shaun ZhouandClaude Opus 5 2026-09-23 15:22:58 +08:00
• 23562448ba fix(detect): 切片统计的口径 Shaun ZhouandClaude Opus 5 2026-09-23 15:13:18 +08:00
• 474b7b50a0 feat(detect): 推理侧改为切片，与固定 160 输入一致 Shaun ZhouandClaude Opus 5 2026-09-23 15:12:35 +08:00
• 2a92275317 feat(detect): 固定 160x160 输入，只用裁剪与镜像填充，禁止缩放 Shaun ZhouandClaude Opus 5 2026-09-23 12:27:57 +08:00
• 1684fddd7f feat(scripts): 重训脚本可跳过重启 ML backend Shaun ZhouandClaude Opus 5 2026-09-22 13:36:25 +08:00
• 3c964abaa2 fix(scripts): HTTP 错误带上响应体 Shaun ZhouandClaude Opus 5 2026-09-22 13:25:41 +08:00
• 744e5fd18a fix(scripts): push_predictions 的分页与令牌过期 Shaun ZhouandClaude Opus 5 2026-09-22 13:17:53 +08:00
• 9ead8c0a6e fix(scripts): push_predictions 支持 JWT 个人访问令牌 Shaun ZhouandClaude Opus 5 2026-09-22 12:13:20 +08:00
• dce2eef0a7 fix(ml-backend): 推理分批，避免 Retrieve Predictions 撑爆显存；新增批量推送脚本 Shaun ZhouandClaude Opus 5 2026-09-22 12:11:06 +08:00
• 8fa7a9d142 fix(ml-backend): 重训后真正切到新权重 Shaun ZhouandClaude Opus 5 2026-09-21 11:03:14 +08:00
• e61fd5887c feat(ml-backend): Start Training 按钮真正触发重训，启动/重训脚本进 scripts/ Shaun ZhouandClaude Opus 5 2026-09-21 11:00:01 +08:00
• 1b8c8d7cac chore: 忽略 ml_backend 的运行时缓存 Shaun ZhouandClaude Opus 5 2026-09-18 17:36:25 +08:00
• a4d099e41d feat(ml-backend): 把训练好的 YOLO 接进 Label Studio 做预标注 Shaun ZhouandClaude Opus 5 2026-09-18 17:36:12 +08:00
• 154677032f Revert "约束预测框互不重叠"（eba533a、9439cbb） Shaun ZhouandClaude Opus 5 2026-09-18 15:34:13 +08:00
• 36fff984b7 refactor(detect): 可视化产出改为按类别分目录，与 generate_boxes 一致 Shaun ZhouandClaude Opus 5 2026-09-18 15:06:13 +08:00
• 9439cbb56a feat(detect): 训练期的验证也约束预测框互不重叠 Shaun ZhouandClaude Opus 5 2026-09-18 11:30:48 +08:00
• eba533a5b2 feat(detect): 约束预测框互不重叠 Shaun ZhouandClaude Opus 5 2026-09-18 11:19:47 +08:00
• 43c8c4a179 fix(detect): 可视化推理自己按批切分，避免 CUDA OOM Shaun ZhouandClaude Opus 5 2026-09-17 21:47:44 +08:00
• 788bc36ab8 feat(detect): 全量推理可视化，预测框与真值框画在一起 Shaun ZhouandClaude Opus 5 2026-09-17 21:47:02 +08:00
• ad50c38f41 chore(detect): 配置指向人工复核后的数据集 data/film_yolo Shaun ZhouandClaude Opus 5 2026-09-17 16:45:14 +08:00
• 59d4b8a9c4 fix(detect): 钳掉 Label Studio 导出坐标的浮点溢出 Shaun ZhouandClaude Opus 5 2026-09-17 16:44:42 +08:00
• 998e6601d4 feat(detect): 把 Label Studio 的 YOLO 导出装配成带 train/val 划分的数据集 Shaun ZhouandClaude Opus 5 2026-09-17 16:42:53 +08:00
• 65bc6b2045 feat(boxes): 新增传统 CV 定位方法（--method cv），不依赖分类模型 Shaun ZhouandClaude Opus 5 2026-09-17 11:03:13 +08:00
• 32e56cb230 feat(boxes): generate_boxes.py 的 --target 默认改为 label Shaun ZhouandClaude Opus 5 2026-09-15 20:31:40 +08:00
• 7dd0f89d18 fix(label-studio): 图片 URL 按实际的 DOCUMENT_ROOT 生成（--document-root） Shaun ZhouandClaude Opus 5 2026-09-15 20:04:04 +08:00
• 891f930de2 refactor: prepare_label_studio.py 的内容改动（补上 a3e08a9 漏掉的部分） Shaun ZhouandClaude Opus 5 2026-09-15 17:26:22 +08:00
• a3e08a9bbc refactor: prepare_yolo_dataset.py 改为 prepare_label_studio.py，产出放到 runs/ Shaun ZhouandClaude Opus 5 2026-09-15 17:26:05 +08:00
• af0897f8c7 refactor(detect): 数据集改为扁平结构、不划分 train/val，Label Studio 文件放到根目录 Shaun ZhouandClaude Opus 5 2026-09-15 17:19:53 +08:00
• 1a43267b55 feat(detect): --label-studio 一步生成可导入 Label Studio 的复核视图 Shaun ZhouandClaude Opus 5 2026-09-15 17:06:40 +08:00
• 7adc5b6332 feat(detect): YOLO 数据集统一为 PNG、按类别重命名并附 classes.txt Shaun ZhouandClaude Opus 5 2026-09-15 16:45:58 +08:00
• cc053a6e9c fix(deps): add label-studio dependency in pyproject.toml xzhou 2026-09-15 08:26:46 +00:00
• 960f0d753a fix(configs): update training epochs in film.yaml and film.yaml for classifier and YOLO models xzhou 2026-09-15 06:49:16 +00:00
• 012b3efc73 feat(detect): 按来源类别处置无框图片（--empty-policy） Shaun ZhouandClaude Opus 5 2026-09-11 16:34:16 +08:00
• d6c6aca325 feat(detect): 默认模型换成 yolo26n Shaun ZhouandClaude Opus 5 2026-09-11 14:34:16 +08:00
• 0e830fe253 fix(config): update data YAML path from boxes_gradcam to boxes_shap Shaun Zhou 2026-09-11 14:29:27 +08:00
• 8dc2fe6a2e feat(detect): 把弱监督框数据集链接到 data/ 并接上 YOLO 训练 Shaun ZhouandClaude Opus 5 2026-09-11 12:05:50 +08:00
• 86c86b105c refactor: 按实际的五步工作流裁掉无关代码 Shaun ZhouandClaude Opus 5 2026-09-11 10:27:20 +08:00
• 57446c2f10 refactor: 合并为单一的框生成入口，gradcam / shap 可选，默认全量 Shaun ZhouandClaude Opus 5 2026-09-10 16:26:56 +08:00
• 9aa87d31a6 feat(localization): SHAP Partition 定位方法，并把每图框数上限默认收紧为 1 Shaun ZhouandClaude Opus 5 2026-09-10 16:10:50 +08:00
• fc7136fddf feat: 用 Grad-CAM 从分类模型生成弱监督框标签 Shaun ZhouandClaude Opus 5 2026-09-10 15:46:30 +08:00
• 33fdf396ee feat: 新增缺陷分类训练与评估链路 Shaun ZhouandClaude Opus 5 2026-09-10 14:58:41 +08:00
• 33d28dcc7c feat(cluster): clusters/ 默认改用硬链接，便于直接在 runs/ 下继续分子类 Shaun ZhouandClaude Opus 5 2026-09-10 12:04:08 +08:00
• f8654f447b feat(export): 改为按 clusters/ 目录导出，并重定位失效的软链 Shaun ZhouandClaude Opus 5 2026-09-10 11:26:02 +08:00
• 9b31473b27 feat(export): 支持把原父目录名作为文件名前缀 Shaun ZhouandClaude Opus 5 2026-09-08 15:26:00 +08:00
• 6e543660ce feat(configs): 新增 resnet101 的自监督聚类配置 Shaun ZhouandClaude Opus 5 2026-09-07 20:51:16 +08:00
• c9e7e4104c fix(configs): update dataset root path in film.yaml xzhou 2026-09-07 12:36:51 +00:00
• 92c88b785c fix(config): --set 拦下不存在的键，并支持续训时重开学习率曲线 Shaun ZhouandClaude Opus 5 2026-09-07 19:30:17 +08:00
• 506fde0290 fix(cluster): 配置层级写错不再静默忽略，并始终打印生效的超参 Shaun ZhouandClaude Opus 5 2026-09-07 19:11:31 +08:00
• 105cb38831 fix(trainer): 检测训练续训以配置为准，并恢复学习率调度进度 Shaun ZhouandClaude Opus 5 2026-09-07 19:06:37 +08:00
• 01cfff9b02 fix(configs): update optimizer settings and add projection parameters in film_simclr_resnet50.yaml xzhou 2026-09-07 06:25:54 +00:00
• 0f5bb4a5c0 feat(cluster): 加 --train-only，自监督训练与聚类可分两步跑 Shaun ZhouandClaude Opus 5 2026-09-07 11:38:59 +08:00
• 5d395b7c25 chore: 数据迁到 data/raw/ 后同步所有路径引用 Shaun ZhouandClaude Opus 5 2026-09-07 11:19:42 +08:00
• 5f31ed0a92 feat(configs): 新增 resnet50 的自监督聚类配置 Shaun ZhouandClaude Opus 5 2026-09-07 11:07:55 +08:00
• f8a52c9382 feat(scripts): 按聚类结果导出数据集 Shaun ZhouandClaude Opus 5 2026-09-06 13:17:53 +08:00
• 7f13ec5d18 fix(configs): update training epochs and optimizer settings in film_simclr_resnet18.yaml xzhou 2026-09-06 05:05:47 +00:00
• 004d7a1ce7 fix(simclr): 换优化器种类时不再加载不兼容的动量缓冲 Shaun ZhouandClaude Opus 5 2026-09-05 11:10:04 +08:00
• b7505a743d fix(simclr): 续训以配置为准，不再被存档里的超参静默覆盖 Shaun ZhouandClaude Opus 5 2026-09-04 21:46:40 +08:00
• 703ba7076b fix(configs): update num_clusters and training parameters in film_simclr_resnet18.yaml xzhou 2026-09-04 13:41:41 +00:00
• 865cdd7210 fix(simclr): 续训时改变总轮数不再让学习率卡在 0 Shaun ZhouandClaude Opus 5 2026-09-04 17:25:45 +08:00
• 6b11128357 fix(configs): update film_simclr_resnet18.yaml with new clustering parameters and optimizer settings xzhou 2026-09-04 09:19:15 +00:00
• 6f00074199 feat(cluster): 新增 resnet18 自监督聚类配置，并修复 --set 的 YAML 标量解析 Shaun ZhouandClaude Opus 5 2026-09-04 15:51:28 +08:00
• 47fb0f1125 fix(configs): 修复 resnet18 训练配置无法构建，并补上对应的评估配置 Shaun ZhouandClaude Opus 5 2026-09-04 15:42:14 +08:00
• e581d9b3d7 fix(backbones): 修复 shufflenet 无法构建，并让换骨架不再踩配置合并的坑 Shaun ZhouandClaude Opus 5 2026-09-04 15:28:49 +08:00
• e4bd794025 feat(cluster): 输出簇分布图，直观展示簇间距离 Shaun ZhouandClaude Opus 5 2026-09-04 15:09:35 +08:00
• 4e6c8aa2ea fix: update work directory in film.yaml and add pearlescent.yaml configuration xzhou 2026-09-04 06:54:06 +00:00
• a92d1125d4 fix: uv 换源至清华源 xzhou 2026-09-04 06:26:16 +00:00
• 4550fdfa8e feat(cluster): add ImageNet pre-trained model section and adjust batch sizes Shaun Zhou 2026-09-03 15:47:37 +08:00
• 345dee50da feat(simclr): 支持断点续训，恢复优化器与调度器状态 Shaun ZhouandClaude Opus 5 2026-09-03 09:34:55 +08:00
• be6173c746 feat(simclr): 接入 SwanLab 可视化，并补上本地看板缺失的依赖 Shaun ZhouandClaude Opus 5 2026-09-03 09:24:39 +08:00
• 7d5b6ec423 docs(readme): 补全 SimCLR 训练与复用权重聚类的完整流程 Shaun ZhouandClaude Opus 5 2026-09-03 09:10:36 +08:00
• 938d62ea53 docs: SimCLR 域内自监督的实测对比与多后端说明 Shaun ZhouandClaude Opus 5 2026-09-02 21:43:35 +08:00
• b6f83fcad0 fix(simclr): 周期性落盘，中途中断不再丢失训练成果 Shaun ZhouandClaude Opus 5 2026-09-02 21:43:35 +08:00
• 141cb907f2 feat: 支持 cuda / mps / cpu 三后端，Mac 上可用 MPS 训练 Shaun ZhouandClaude Opus 5 2026-09-02 21:43:35 +08:00
• 4fc0fa5353 docs: 无监督聚类的数学说明、mermaid 图与 README Shaun ZhouandClaude Opus 5 2026-09-02 17:55:48 +08:00
• d6adea5b05 test: 单元测试与端到端冒烟 Shaun ZhouandClaude Opus 5 2026-09-02 17:55:47 +08:00
• f433e5cf59 feat: 未分类样本的无监督聚类（多 stage 特征 + KMeans，可选 SimCLR） Shaun ZhouandClaude Opus 5 2026-09-02 17:55:47 +08:00
• 4b1147ed8f feat(scripts): 数据转换、ONNX 导出与推理性能分析 Shaun ZhouandClaude Opus 5 2026-09-02 17:55:20 +08:00
• b0d0dfffbb feat: 训练与评估入口及 yaml 配置 Shaun ZhouandClaude Opus 5 2026-09-02 17:55:20 +08:00
• ef64e5bf77 feat(engine): 训练评估循环、CenterNet 损失与指标 Shaun ZhouandClaude Opus 5 2026-09-02 17:55:20 +08:00
• af22dc639c feat(datasets): YOLO 风格标注数据集与几何/像素增强 Shaun ZhouandClaude Opus 5 2026-09-02 17:55:20 +08:00
• 3a36aad6c3 feat(models): 纯卷积缺陷分类定位模型 Shaun ZhouandClaude Opus 5 2026-09-02 17:54:55 +08:00
• f4eff8beee feat(utils): 配置继承、注册表、日志与 Nvidia GPU 守卫 Shaun ZhouandClaude Opus 5 2026-09-02 17:54:55 +08:00
• 0645fc19f7 chore: 初始化 uv 项目配置与忽略规则 Shaun ZhouandClaude Opus 5 2026-09-02 17:54:55 +08:00
• c0e7c23aef Initial commit xzhou 2026-09-01 16:48:33 +08:00