Skip to content

关于训练Train Advantage Estimator #12

@caokaifa

Description

@caokaifa

感谢分享好的作品,在训练优势估计时有几个问题:
1.代码运行错误,看到config.py中class DataConfig没有episodes,但是452行需要,这个地方是不是错误还是我搞错了?
2.

In src/openpi/training/config.py, find ADVANTAGE_TORCH_PI06_FLATTEN_FOLD or ADVANTAGE_TORCH_KAI0_FLATTEN_FOLD:

TrainConfig(
name="ADVANTAGE_TORCH_KAI0_FLATTEN_FOLD", # or ADVANTAGE_TORCH_PI06_FLATTEN_FOLD
data=LerobotAgilexDataConfig(
repo_id="<your_labeled_dataset_path>", # <-- update this
assets=AssetsConfig(
assets_dir="<your_labeled_dataset_path>/assets", # <-- update this
asset_id="<your_dataset_name>", # <-- update this
),
),
pytorch_weight_path="<path_to_pi05_base_checkpoint>", # <-- update this
...
)
Image这个assets_dir与asset_id不是配置找到归一化信息的吗?配置里也配置了不需要归一化,是说数据需要归一化后配置这个吗?

AWBC training (Stage 3) without running Stage 0–2. It is available in both the Hugging Face and ModelScope dataset repos. After downloading (e.g. via scripts/download_dataset.py), set the AWBC config repo_id to the local path (e.g. <repo_root>/data/Task_A/advantage) and run training.
Image
这个说Stage0-2,但是我下载你们的数据advantage里面没有data_PI06_100000/ or data_KAI0_100000/,我理解错了吗?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions