Skip to content

About

Merging tool for W&B runs on offline and time limited clusters

Resources

Stars

1 star

Watchers

0 watching

Forks

Latest commit

 

History

1 Commit

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 

Repository files navigation

wandb-merger

Inspect, merge, and sync offline W&B (.wandb) run files.

Useful when training is split across multiple runs (crash/resume, multi-stage) and you want a single unified run on W&B.

Install

pip install git+https://github.com/MertArslanoglu/wandb-merger.git

CLI Commands

Command Description
wandb-inspect Inspect a .wandb file offline
wandb-merge Merge two .wandb runs into one
wandb-merge-dir Auto-discover & merge all runs in a directory
wandb-merge-sync Merge + sync to W&B cloud in one step

Quick start

# List all runs in a saved directory
wandb-merge-dir --saved-dir saved/ --list-runs

# Merge and sync everything
wandb-merge-sync --saved-dir saved/ --entity myteam --project MyProject --name "full_run"

More examples

# Inspect a single run
wandb-inspect /path/to/run-dir/ --show-config --show-keys

# Merge two runs
wandb-merge --first run1/ --second run2/ --output merged/

# Merge a directory, filter short runs, custom validation prefix
wandb-merge-dir --saved-dir saved/ --output merged/ --min-runtime 60

Features

  • Auto-offset: steps, epochs, runtime, and timestamps are automatically continued across runs
  • Auto-detect continuation: if a resumed run already starts from the correct epoch, no duplicate offset is applied
  • Validation injection: validation runs are injected as labeled metrics (e.g. SemanticKITTI/validation/mIoU_epoch)
  • Crash/resume support: handles folders with multiple offline-run-* dirs from crash+resume

Python API

from wandb_merger import (
    read_wandb_file, find_wandb_file, history_to_dataframe,
    merge_wandb_files, merge_multiple_runs,
    discover_runs, split_by_type,
)

About

Merging tool for W&B runs on offline and time limited clusters

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages