low-high-reference / README.md
Ouzhang's picture
Document benchmark setup commands
076b0f3 verified
|
Raw
History Blame Contribute Delete
6.85 kB
# Reference Directory
git clone https://github.com/PKU-YuanGroup/Helios.git
git clone https://github.com/NVlabs/LongLive.git
git clone git@github.com:bingreeky/MemGen.git
这个目录用于存放项目设计、实现和训练过程中会反复参考的外部资料。
它不是运行时必须的源码目录,而是研究与开发参考层。
## 目录定位
`reference/` 主要存放以下几类内容:
- 论文 PDF
- 论文配套笔记
- 方案草稿
- 外部开源项目的结构化阅读记录
- 和当前项目直接相关的训练/规划/critic/memory 参考材料
它的作用不是“被 import”,而是帮助回答这些问题:
- 当前系统应该如何拆成 `planner / edit / critic / memory`
- 哪些训练阶段适合先做监督、后做偏好、再做 RL / GRPO
- 哪些工作可以作为 `skill / memory / reflection` 的设计参考
- 在汇报和写文档时,相关设计依据来自哪里
## 当前内容类型
目前这个目录下的内容大致包括:
- `SPIRAL- ... .pdf`
- 用来参考闭环 `planner -> world model -> critic -> memory` 的训练故事
- `HGM/`
- `HyperAgents/`
- `MemGen/`
- `OpenSpace/`
- `ViMax/`
- `others/`
这些子目录一般表示:
- 某篇论文的阅读材料
- 某个参考项目的摘录或整理
- 某类方法的补充资料
## 在当前项目中的具体用途
### 1. Planner 设计
参考:
- 结构化 planning
- style family 选择
- carrier / preserve / rewrite 的拆解方式
当前对应代码:
- [src/planner/core/video_planner.py](/home/featurize/low-high-new/src/planner/core/video_planner.py)
### 2. Critic 设计
参考:
- evaluator 评分维度
- reflection policy
- pairwise reward / preference
- 闭环修正机制
当前对应代码:
- [src/critic/evaluator.py](/home/featurize/low-high-new/src/critic/evaluator.py)
- [src/critic/reflector.py](/home/featurize/low-high-new/src/critic/reflector.py)
### 3. Memory / Skill 设计
参考:
- 显式记忆
- skill 积累
- replay 驱动更新
- 成功/失败经验抽象
当前对应代码:
- [src/memory/store.py](/home/featurize/low-high-new/src/memory/store.py)
- [src/memory/retriever.py](/home/featurize/low-high-new/src/memory/retriever.py)
- [src/memory/evolver.py](/home/featurize/low-high-new/src/memory/evolver.py)
### 4. Edit-model 训练设计
参考:
- Ditto 风格的 `low/high/prompt` 监督训练
- SPIRAL 风格的闭环优化
- preference / GRPO 的引入时机
当前对应代码:
- [scripts/training/edit-model/train.py](/home/featurize/low-high-new/scripts/training/edit-model/train.py)
## 为什么这个目录需要在 Git 里可见
虽然 `reference/` 里的大部分内容不是运行时必需代码,但它对项目协作很重要:
- 新加入的人可以快速知道设计参考来自哪里
- 写 README、汇报、实验计划时可以快速对齐依据
- 避免“本地有一堆参考资料,但 Git 仓库里完全看不见”
因此当前仓库采用的策略是:
- `reference/` 目录内容默认不整体纳入版本管理
- 但至少保留这个 `README.md`,说明目录用途和内容范围
## 配置 benchmarks 目录
`benchmarks/` 里存放评测相关论文、README 摘录、外部 benchmark 代码阅读记录,以及我们整理的落地脚本。它目前通过 Hugging Face dataset repo 保存,不建议把大 PDF、外部代码和缓存文件直接塞进主仓库。
### 方式一:随父仓库 submodule 拉取
如果你是在完整的 `low-high-new` 仓库里工作,优先用这种方式:
```bash
cd /path/to/low-high-new
git pull
git submodule update --init --recursive reference
```
如果只想强制更新 `reference` 到父仓库记录的版本:
```bash
cd /path/to/low-high-new
git submodule sync -- reference
git submodule update --init --recursive --force reference
```
### 方式二:直接从 Hugging Face 下载 reference
如果当前机器没有完整初始化 submodule,或者只想恢复 `reference/benchmarks` 内容,可以直接用 HF 下载:
```bash
cd /path/to/low-high-new
hf download Ouzhang/low-high-reference \
--repo-type dataset \
--local-dir reference \
--include 'benchmarks/**' \
--exclude '**/.DS_Store' \
--exclude '**/__pycache__/**'
```
如果只需要视频编辑评测资料和脚本:
```bash
cd /path/to/low-high-new
hf download Ouzhang/low-high-reference \
--repo-type dataset \
--local-dir reference \
--include 'benchmarks/edit/**' \
--exclude '**/.DS_Store' \
--exclude '**/__pycache__/**'
```
### 当前 edit benchmark 重点文件
```bash
reference/benchmarks/edit/评测表格.md
reference/benchmarks/edit/traditional_eval_notes.md
reference/benchmarks/edit/build_six_method_manifest.py
reference/benchmarks/edit/run_traditional_metrics.py
```
生成我们当前 6 组 `val20/val100` manifest:
```bash
cd /path/to/low-high-new
python3 reference/benchmarks/edit/build_six_method_manifest.py \
--repo-root . \
--output-dir out/edit_model_face_stage1/traditional_eval_manifests
```
先跑不需要下载模型的传统指标:
```bash
cd /path/to/low-high-new
python3 reference/benchmarks/edit/run_traditional_metrics.py \
--manifest out/edit_model_face_stage1/traditional_eval_manifests/val20.jsonl \
--output out/edit_model_face_stage1/traditional_metrics_val20.jsonl \
--frames-per-video 16
python3 reference/benchmarks/edit/run_traditional_metrics.py \
--manifest out/edit_model_face_stage1/traditional_eval_manifests/val100.jsonl \
--output out/edit_model_face_stage1/traditional_metrics_val100.jsonl \
--frames-per-video 16
```
如果已经下载好 CLIP,可以额外跑 CLIP 文本对齐和跨帧一致性:
```bash
cd /path/to/low-high-new
hf download openai/clip-vit-large-patch14 \
--local-dir models/openai_clip-vit-large-patch14
python3 reference/benchmarks/edit/run_traditional_metrics.py \
--manifest out/edit_model_face_stage1/traditional_eval_manifests/val100.jsonl \
--output out/edit_model_face_stage1/traditional_metrics_val100_clip.jsonl \
--frames-per-video 16 \
--clip-model-dir models/openai_clip-vit-large-patch14 \
--device cuda:0
```
注意:`benchmarks/edit/code/` 下如果放了外部 benchmark 原始仓库,通常还需要按各自 README 单独安装依赖;本 README 只负责说明如何恢复资料和运行我们整理的轻量脚本。
## 使用建议
- 新增参考资料时,优先放进已有子目录或按主题新建子目录
- 如果某份资料对当前实现有直接影响,建议在提交说明或设计文档中顺手写明
- 不建议把过大的无关临时下载内容混进这里
## 总结
`reference/` 是本项目的“研究参考层”,不是运行时依赖层。
它的价值在于:
- 对齐设计依据
- 追踪方法来源
- 帮助理解为什么当前系统被设计成现在这样