deep-learning-wandb

Public Weights & Biases integration layer for deep-learning-core.

deep-learning-wandb adds a W&B callback and scaffold integration on top of deep-learning-core. It keeps tracking-specific logic outside the core framework while still allowing users to install it through deep-learning-core[wandb].

Current release: deep-learning-wandb==0.0.17. Requires deep-learning-core>=0.1.8,<0.2.

What's New in 0.0.17?

• W&B logs omit invalid scalar metrics and finish runs with explicit terminal status
• runtime extension registration and scaffold setup work with dl-core 0.1.8
• the development PyTorch requirement is torch>2.3 without an upper cap

Previous versions are recorded in the release history.

Install

Install from PyPI through the core extra:

pip install "deep-learning-core[wandb]"

Install the package directly:

pip install deep-learning-wandb

Install in a uv project:

uv add "deep-learning-core[wandb]"

Scope

• W&B callback registration for deep-learning-core
• Experiment scaffold integration through dl-init --with-wandb
• W&B-ready config defaults for generated experiment repositories

Out Of Scope

• Generic trainer, dataset, and metric abstractions
• Azure execution or storage logic
• Company-specific W&B entities, projects, or secrets

Quick Start

Install it through the deep-learning-core extra:

uv add "deep-learning-core[wandb]"

Then scaffold a W&B-ready experiment repository:

uv run dl-init --name my-exp --with-wandb

The installed package registers the wandb callback through dl-core's runtime extension entry points. The generated package also imports dl_wandb. It also ignores .env, other local environment files, and wandb/, while keeping .env.example available as the credential template.

Concrete experiment flow:

uv init
uv add deep-learning-wandb
uv run dl-init --root-dir . --with-wandb
uv run dl-run --config configs/base.yaml
uv run dl-sweep experiments/lr_sweep.yaml

The W&B project defaults to the repository root name unless tracking.experiment_name overrides it. The sweep file name becomes the W&B run group unless tracking.sweep_name overrides it.

What You Get

• the wandb callback for local training runs
• epoch, RL episode, algorithm-update, and evaluation metric logging, with RL series indexed by the environment-transition global_step
• epoch metrics aligned with the trainer and history.json epoch index
• nonzero W&B exit codes for failed and interrupted runs, with the exact terminal state retained in the run summary
• dl-init --with-wandb scaffold support
• generated W&B callback defaults and .env.example

Companion Packages

• dl-core
• dl-azure
• dl-mlflow

Documentation

• Documentation Index
• GitHub Repository

License

MIT. See LICENSE.