Skip to content

config_reference.md

Latest commit

 

History

History
83 lines (65 loc) · 3.17 KB

config_reference.md

File metadata and controls

83 lines (65 loc) · 3.17 KB
Raw

Configuration Reference

This document explains the YAML configuration structure used by LayoutVLM-Lab. Example file: config/paddle_layoutv2.yaml.

Output structure

Results are written to output.output_dir/<experiment.name>/:

  • Markdown outputs are saved directly in this folder
  • json/ structured results
  • imgs/ cropped image blocks (optional)

Global sections

experiment

Key Type Default Description
name string "default" Run name used to create the output folder.

output

Key Type Default Description
output_dir string PROJECT_ROOT/results Base output directory for all runs.
save_images bool true Save cropped image blocks for labels mapped to image.

input

Key Type Default Description
image_root string PROJECT_ROOT/images Directory containing input images.
filter_processed bool true Skip images that already have JSON results in json/.

pipeline

Key Type Default Description
vlm_workers int 20 Number of VLM worker threads.
vlm_queue_max_size int 50 Max VLM queue size (recommend ~2x workers).
block_workers int 0 Block-level VLM thread pool size per image. 0 uses auto (min(8, cpu_count)).

Layout section

layout

Key Type Default Description
type string required Layout backend id (e.g. paddle).
config object {} Backend-specific settings passed through.
label_mapping object {} Mapping from layout labels to metric types.

layout.config (common fields)

  • model_name: layout model name (e.g. PP-DocLayoutV2)
  • model_dir: model weights directory
  • device_list: GPU ids, comma-separated (e.g. "0,1")
  • nums_per_gpu: processes per GPU
  • merge_blocks_before_md: merge blocks before Markdown export

label_mapping

  • ignore: labels ignored during Markdown export
  • image: labels skipped by VLM; crops saved to imgs/ if save_images is true
  • table: labels processed with VLM table prompt
  • formula: labels processed with VLM formula prompt
  • others default to ocr

VLM section

vlm

Key Type Default Description
type string required VLM backend id (e.g. openai).
config object {} Backend-specific settings passed through.
metrics list [] Enabled metric types; only these blocks are sent to VLM.

vlm.config (OpenAI-compatible)

  • model_name
  • api_key / base_url (can be ${OPENAI_API_KEY} / ${OPENAI_API_BASE_URL})
  • max_tokens, temperature, timeout
  • text_before_image

Environment variables

Config.from_yaml resolves ${VAR} placeholders using environment variables from .env.

Backward compatibility

If experiment.output_dir is present and output.output_dir is missing, the loader will still use it.

If both are omitted, it falls back to PROJECT_ROOT/results/<experiment.name> (with experiment.name defaulting to default), and the json/ and imgs/ subfolders are created automatically.