General Conventions#

This page defines the unified conventions used throughout the API documentation, facilitating consistent interpretation of parameter semantics and usage across different sections.

Parameter Description Template#

This documentation adopts the following parameter table structure by default:

Parameter

Type

Default

Description

xxx

str/int/flag/path/key=value

...

Parameter purpose, value constraints, and notes

Notes:

  • Parameter: Command argument or configuration field name (original identifier preserved).

  • Type: Expressed according to existing documentation conventions, e.g., path, flag, key=value pair.

  • Default: Only default values explicitly stated in the existing documentation are recorded.

  • Description: Priority is given to describing the scope of effect and relationships with other parameters.

Naming and Path Conventions#

Configuration Paths#

  • CONFIG in training/evaluation commands refers to the configuration file path (e.g., configs/.../*.py).

  • WORK_DIR refers to the output directory for runtime artifacts (logs, checkpoints, configuration snapshots, etc.).

  • CKPT_PATH refers to the checkpoint file used for evaluation or inference.

Script Parameters#

  • Training entry point: bash scripts/train.sh [CONFIG] [WORK_DIR] [additional arguments...]

  • Evaluation entry point: bash scripts/eval.sh [CONFIG] [CKPT_PATH] [additional arguments...]

  • Additional arguments are generally passed through to train.py / eval.py (e.g., --cfg-options).

Environment Variables#

Distributed training and evaluation rely on the MLP_* environment variables:

  • MLP_WORKER_GPU

  • MLP_WORKER_NUM

  • MLP_ROLE_INDEX

  • MLP_WORKER_0_HOST

  • MLP_WORKER_0_PORT

Commonly used environment variables for experiment logging:

  • WANDB_PROJECT

  • WANDB_ENTITY

  • WANDB_MODE

Input/Output Field Conventions (Documentation Level)#

The following are frequently referenced field groups in the documentation, provided for cross-page lookup:

  • Observation-related: states, observation.state, observation.eepose, images, img_masks

  • Language-related: lang_tokens, lang_masks, prompt, task_description

  • Action-related: actions, action_masks, action_dim, ori_action_dim

  • Top-level configuration keys: model, train_dataloader, runner, inference, eval

Note: This page only consolidates fields that explicitly appear in the existing Chinese documentation and does not extend to undisclosed implementation details.

Minimal Example#

# 训练
bash scripts/train.sh configs/xxx.py work_dirs/xxx

# 评估
bash scripts/eval.sh configs/xxx.py work_dirs/xxx/checkpoint_step_10000.pt

Common Error References#