kaggle-wandb-sync

A CLI tool to sync Weights & Biases offline runs from Kaggle Notebooks to W&B cloud — fully automated via GitHub Actions.

Why?

Kaggle Notebooks run in an isolated environment with internet access disabled for competition submissions. This means you can't push W&B metrics in real time. kaggle-wandb-sync solves this by:

1. Running your notebook with WANDB_MODE=offline (logs saved locally on Kaggle)
2. Downloading the output via kaggle kernels output
3. Syncing the offline runs to W&B cloud with wandb sync

Installation

pip install kaggle-wandb-sync

Prerequisites: Kaggle API credentials (~/.kaggle/kaggle.json) and a W&B API key (WANDB_API_KEY env var, or run wandb login once to save credentials to ~/.netrc).

Quick Start

All-in-one command

# Set your W&B API key
export WANDB_API_KEY=your_api_key

# Push notebook, wait for completion, download output, sync to W&B
kaggle-wandb-sync run my-notebook/

Step by step

kaggle-wandb-sync push   my-notebook/                      # push (with 409 protection)
kaggle-wandb-sync poll   username/my-notebook              # wait for COMPLETE
kaggle-wandb-sync output username/my-notebook              # download output
kaggle-wandb-sync sync   ./kaggle_output                   # wandb sync

Notebook Setup

Add these lines before importing wandb in your Kaggle Notebook:

import os
os.environ['WANDB_MODE'] = 'offline'   # must be set before import
os.environ['WANDB_PROJECT'] = 'my-project'

import wandb
wandb.init()
# ... your training code ...
wandb.log({"loss": 0.1, "accuracy": 0.95})
wandb.finish()

Important: Set WANDB_MODE=offline before import wandb, not after.

GitHub Actions Integration

Add this workflow to your Kaggle repo (.github/workflows/kaggle-wandb-sync.yml):

name: Kaggle W&B Sync

on:
  workflow_dispatch:
    inputs:
      notebook_dir:
        description: "Notebook directory (e.g. my-competition)"
        required: true
      kernel_id:
        description: "Kernel ID (e.g. username/my-competition-baseline)"
        required: true

jobs:
  sync:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4

      - uses: actions/setup-python@v5
        with:
          python-version: "3.12"

      - name: Install kaggle-wandb-sync
        run: pip install kaggle-wandb-sync

      - name: Set up Kaggle credentials
        run: |
          mkdir -p ~/.kaggle
          echo '${{ secrets.KAGGLE_API_TOKEN }}' > ~/.kaggle/kaggle.json
          chmod 600 ~/.kaggle/kaggle.json

      - name: Run pipeline
        env:
          WANDB_API_KEY: ${{ secrets.WANDB_API_KEY }}
        run: |
          kaggle-wandb-sync run ${{ inputs.notebook_dir }} \
            --kernel-id ${{ inputs.kernel_id }}

Required secrets: KAGGLE_API_TOKEN (JSON content of ~/.kaggle/kaggle.json) and WANDB_API_KEY.

Commands

kaggle-wandb-sync output

Download output files from a completed Kaggle kernel.

kaggle-wandb-sync output [KERNEL_ID] [OPTIONS]

kaggle-wandb-sync poll

Poll a Kaggle kernel until it reaches COMPLETE, ERROR, or CANCEL.

kaggle-wandb-sync poll [KERNEL_ID] [OPTIONS]

kaggle-wandb-sync push

Push a Kaggle Notebook to Kaggle.

kaggle-wandb-sync push [DIRECTORY] [OPTIONS]

kaggle-wandb-sync run

Run the full pipeline: push → poll → output → wandb sync → wait for submission → record LB score.

kaggle-wandb-sync run [DIRECTORY] [OPTIONS]

kaggle-wandb-sync score

Log Kaggle submission scores to a W&B run.

kaggle-wandb-sync score [RUN_ID] [OPTIONS]

kaggle-wandb-sync sync

Sync W&B offline runs found in OUTPUT_DIR to W&B cloud.

kaggle-wandb-sync sync [OUTPUT_DIR] [OPTIONS]

Command notes

Behaviour that is not visible in the option tables above:

• push waits for any currently running kernel to finish before pushing, which prevents 409 Conflict errors.

• poll exits with code 1 if the kernel finishes with ERROR or CANCEL. Since v0.1.5 it also downloads the kernel log on those outcomes and prints stdout plus the last 30 stderr lines, so you can diagnose a failure without opening the Kaggle UI.

• sync finds every offline-run-* directory under the output dir and runs wandb sync on each.

• score takes a full run URL, an entity/project/id path, or a bare id with --project:

  kaggle-wandb-sync score https://wandb.ai/me/my-proj/runs/abc123 --score 0.127 --rank 200
  

The section above Command notes is generated from the Click definitions by scripts/gen_commands_doc.py and rewritten by CI on every push. Put anything hand-written here, below <!-- commands:end -->, or it will be overwritten.

Known Issues

• Windows encoding: Prefix commands with PYTHONUTF8=1 if you see encoding errors on Windows.

• Windows PATH (Microsoft Store Python): If kaggle-wandb-sync: command not found in Git Bash, add the Scripts directory to your PATH:

  # Add to ~/.bashrc
  export PATH="$PATH:/c/Users/<your-username>/AppData/Local/Packages/PythonSoftwareFoundation.Python.3.12_qbz5n2kfra8p0/LocalCache/local-packages/Python312/Scripts"
  

• Git Bash path format (fixed in v0.1.2): Git Bash converts paths like C:/Users/... to /c/Users/..., which Python cannot resolve. As of v0.1.2, all path arguments are automatically converted to Windows format.

License

MIT