instant-fold-assets
收藏资源简介:
INSTANT-FOLD Assets 数据集是 INSTANT-FOLD 项目的公共发布资产集合,主要用于布料模拟、轨迹生成、模型预训练和基准评估。数据集包含多个资产包,每个包针对特定用途:eval 包用于普通公共模拟评估,包含检查点、演示轨迹和网格数据;data-gen 包提供用于本地公共轨迹生成的 360 个网格库;benchmark-eval 包支持基准等效的精确初始化回放评估;benchmark-repro 包提供紧凑资产,用于复现发布的相同初始化基准表格;pretrain-smoke 包为预训练提供小型可运行的冒烟测试语料;pretrain-full 包则包含完整的上下文 1-8 公共 LoRA 预训练语料库,采用 H5 格式。数据集适用于布料物理模拟、强化学习策略评估、模型预训练以及研究复现等任务。
The INSTANT-FOLD Assets dataset is a public release asset collection for the INSTANT-FOLD project, primarily used for cloth simulation, trajectory generation, model pre-training, and benchmark evaluation. The dataset includes multiple asset packages, each tailored for specific purposes: the eval package is for general public simulation evaluation, containing checkpoints, demonstration trajectories, and mesh data; the data-gen package provides a library of 360 meshes for local public trajectory generation; the benchmark-eval package supports precise initialization replay evaluation equivalent to benchmarks; the benchmark-repro package offers compact assets for reproducing the same initialization benchmark tables as published; the pretrain-smoke package provides a small, runnable smoke test corpus for pre-training; and the pretrain-full package contains the full context 1-8 public LoRA pre-training corpus in H5 format. The dataset is suitable for tasks such as cloth physics simulation, reinforcement learning policy evaluation, model pre-training, and research reproduction.
INSTANT-FOLD Assets 数据集概述
数据集基本信息
- 许可证:MIT
- 用途:存储 INSTANT-FOLD 项目的公开发布资产
数据包(Bundles)概览
| 数据包 | 用途 | 典型命令 |
|---|---|---|
encoder |
已发布的预训练编码器检查点 | instant-fold assets download --bundle encoder |
eval |
普通公开仿真评估 | instant-fold assets download --bundle eval |
data-gen |
已发布的 360 网格库,用于本地公开轨迹生成 | instant-fold assets download --bundle data-gen |
benchmark-eval |
基准等价评估的精确初始化重放资产 | instant-fold assets download --bundle benchmark-eval |
benchmark-repro |
复现已发布的相同初始化基准表的紧凑资产 | instant-fold assets download --bundle benchmark-repro |
train-smoke |
可直接运行的公开训练数据微型语料库 | instant-fold assets download --bundle train-smoke |
train-full |
完整的 ctx1-8 公开共享训练语料库 | instant-fold assets download --bundle train-full |
数据包内容详解
encoder
checkpoints/pretrained_lora.pth:用于表示可视化与公开策略训练的预训练编码器检查点
eval
checkpoints/8-ctx-policy.pthdata/eval_demo_trajectories.tardata/eval_mesh_ids.txtdata/cloth3d_mesh_tops_eval_subset.tar.gz
benchmark-eval
benchmark_replay/replay_manifest.jsonbenchmark_replay/eval_state_cache.tar.gz:用于实现与已发布基准等价的精确初始化重放
benchmark-repro
benchmark_repro/reference_features_38400.npzbenchmark_repro/reference_metric_rows_38400.csvbenchmark_repro/language_conditioned.tar.gzbenchmark_repro/demo_conditioned_8_ctx_policy.tar.gz- 用于使用紧凑公开缓存复现以下设置的相同初始化基准表:
language_conditioned、demo_conditioned_8_ctx_policy、oracle
data-gen
data/cloth3d_mesh_tops_full.tar.gz:360 网格库,安装后位于data/meshes/tops/,用于本地轨迹生成
train-smoke
train/train_smoke.tar:可直接运行的公开训练数据微型语料库
train-full
train/train_full_metadata.tartrain/train_ctx1_8_way1.tar至train/train_ctx1_8_way8.tar(共 8 部分)- 完整的 ctx1-8 H5 语料库,同时用于编码器预训练和策略训练
使用建议
- 普通用户:仅需
eval数据包 - 基准复现:需要同时下载
eval和benchmark-eval数据包 - 论文表格复现:需要使用
benchmark-repro数据包 - 本地数据生成:使用
data-gen数据包 - 模型训练:根据需求选择
train-smoke(轻量测试)或train-full(完整训练)




