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==0.0.14. Requires deep-learning-core>=0.0.34,<0.1.

What's New in 0.0.14?

• W&B now implements the public RL callback hooks on_episode_end(), on_update_end(), and on_evaluation_end()
• custom callbacks can subclass the integration without relying on private underscore methods

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 generated experiment package will import dl_wandb automatically so the wandb callback registers at runtime. 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
• 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.