I built KaggleFleet to simplify collaborative Kaggle GPU workflows
Working on ML experiments with multiple collaborators often means dealing with multiple notebook runs, Kaggle accounts, and output files. I wanted to make that process less repetitive, so I built KaggleFleet . The idea Instead of manually switching between Kaggle accounts and tracking notebook execution separately, KaggleFleet gives you one local interface to manage the workflow. You can keep…
Collaborating on machine learning projects with multiple researchers can involve numerous notebook runs, account switches, and disparate output files. To streamline this process, I developed KaggleFleet, a tool designed to simplify collaborative Kaggle GPU workflows. Rather than manually toggling between accounts and tracking notebook execution separately, KaggleFleet offers a unified local interface to manage the entire workflow.
Users can retain their notebooks within a local project while selecting an account, choosing an accelerator, pushing the notebook, and monitoring its execution.
For research teams working together, KaggleFleet accommodates multiple, independently authenticated co-author accounts and enables simultaneous notebook execution across each account. Key features include:
1. Local-first notebook workflow: Keep your notebooks within your project and utilize your preferred editor.
2. Multiple accounts: Configure distinct co-author accounts independently via Kaggle's browser authentication flow.
3. Parallel execution: Launch experiments using different accounts concurrently, without manual account switching.
4. Centralized monitoring: Employ a local web panel to view accounts, active jobs, statuses, and logs.
5. Output collection: Download completed kernel outputs into your local project for easy access.
6. AI-agent integration: Commands generate JSON output and meaningful exit codes, facilitating integration with coding agents and scripts.
To begin using KaggleFleet, install it with `pipx install git+https://github.com/towhaEL/KaggleFleet.git` or `pip install kaggle-fleet`. Then, execute `kfleet doctor` to verify system compatibility, `kfleet account login --label "Me"` to authenticate your account, `kfleet notebook list` to view your notebooks, `kfleet run push train.ipynb --watch` to push and monitor a notebook, and `kfleet serve --open` to launch the notebook in your default browser.
KaggleFleet is currently seeking early adopters, particularly ML researchers and those engaged in collaborative Kaggle projects, to gather feedback on any remaining pain points and additional features needed to make the tool truly invaluable. Visit the GitHub repository at https://github.com/towhaEL/KaggleFleet for more information.
Rewritten by urgent.news; may contain errors — check the original.
Source: dev.to