CLI Reference

The unified AutoPipeline entry point is:

autopipeline

If you have not installed the wrapper command, you can call the module directly:

python -m src.cli.autopipeline

Subcommands

• annotation
• eval
• train-pairs

note

The current implementation hardcodes many defaults under /data/open_edit/.... If your checkout lives elsewhere, pass --pipeline-config-path, --user-config, --save-path, and related paths explicitly.

annotation

Purpose: run a human-centric or object-centric pipeline and score candidate edited images with structured metrics.

autopipeline annotation \
  --edit-task <task> \
  --pipeline-config-path <pipeline-yaml> \
  [--max-workers 4] \
  [--save-path /data/open_edit/data/c_annotated_group_data] \
  [--user-config /data/open_edit/configs/pipelines/user_config.yaml] \
  [--candidate-pool-dir /data/open_edit/configs/datasets/candidate_pools]

Parameters:

Parameter	Required	Description
--edit-task	Yes	Edit task name. The CLI normalizes it into lowercase underscore form.
--pipeline-config-path	Yes	Absolute path to the pipeline YAML.
--max-workers	No	Worker parallelism.
--save-path	No	Output directory for results and cache.
--user-config	No	User config YAML.
--candidate-pool-dir	No	Directory containing candidate pool JSON files.

eval

Purpose: run vlm-as-a-judge and produce pairwise winners on a benchmark.

autopipeline eval \
  --bmk <benchmark-key> \
  --pipeline-config-path <pipeline-yaml> \
  [--max-workers 4] \
  [--save-path /data/open_edit/data/reward_eval_results] \
  [--user-config /data/open_edit/configs/pipelines/user_config.yaml] \
  [--bmk-config /data/open_edit/configs/datasets/bmk.json] \
  [--openedit-metadata-file metadata.jsonl]

Parameters:

Parameter	Required	Description
--bmk	Yes	Benchmark key defined in bmk.json.
--pipeline-config-path	Yes	Judge pipeline YAML.
--max-workers	No	Worker parallelism.
--save-path	No	Result directory.
--user-config	No	User config YAML.
--bmk-config	No	Benchmark config file.
--openedit-metadata-file	No	Metadata filename used only for openedit evaluation.

train-pairs

Purpose: convert grouped results into preference-training data.

autopipeline train-pairs \
  --tasks <task1,task2,...> \
  [--prompts-num 1500] \
  [--prefix ""] \
  [--input-dir /data/open_edit/data/c_annotated_group_data] \
  [--output-dir /data/open_edit/data/d_train_data] \
  [--mode auto] \
  [--filt-out-strategy three_tiers] \
  [--thresholds-config-file /data/open_edit/configs/pipelines/data_construction_configs.json]

Parameters:

Parameter	Required	Description
--tasks	Yes	Comma-separated task names.
--prompts-num	No	Maximum number of prompt groups per task.
--prefix	No	Optional output subdirectory prefix.
--input-dir	No	Input directory for grouped results.
--output-dir	No	Output directory for train pairs.
--mode	No	auto, group, or judge.
--filt-out-strategy	No	head_tail or three_tiers.
--thresholds-config-file	No	Threshold config for group mode.

Two common invocation styles

User-facing CLI

autopipeline annotation ...
autopipeline eval ...
autopipeline train-pairs ...

Module entry point

python -m src.cli.autopipeline annotation ...
python -m src.cli.autopipeline eval ...
python -m src.cli.autopipeline train-pairs ...