Experiment tracking with Weights & Biases (W&B) in SMLE

A step by step guide on how to integrate Weights & Biases (W&B) into an SMLE-based project in order to track machine learning experiments (metrics, hyperparameters, models, and training logs).

Note

This guide assumes that you have an active W&B account at https://wandb.ai.

1. Obtain your WandB API key

Before you can log experiments from SMLE to W&B, you must create a W&B account and obtain a personal API key associated with that account. This key is used to authenticate all requests from your code to the W&B backend.

  1. Loggin in https://wandb.ai and open your user Settings page from the account menu.

  2. Locate the API Keys section and create or copy your personal API key, which will be required to authenticate the W&B client from your code.

⚠️ Security & WandB Configuration

Warning

When using the wandb section for remote logging, the API key is read from smle.yaml. To avoid exposing credentials, do not commit smle.yaml or log files with real keys to any public repository.

It is a good practice to:

  • Add smle.yaml and *.log to .gitignore.

  • Remove the wandb section entirely if remote logging is not required.

2. Configuring SMLE: the smle.yaml file

SMLE uses a smle.yaml file to centralize project configuration (training, logging, etc.). To enable W&B, add a dedicated section.

Minimal example of smle.yaml with W&B support:

wandb:
     entity: your_wandb_account
     key: your_wandb_key

3. Running an experiment

Once the smle.yaml file has been configured, including the wandb section, you can start an experiment by running your SMLE entrypoint script as a standard Python program:

python main.py

During execution, the script reads the configuration from smle.yaml, initializes the W&B client, and sends configuration data and training metrics to your W&B project. When the run completes, you can open https://wandb.ai, navigate to the configured project, and inspect the dashboards with loss and accuracy curves, the stored experiment configurations, and any saved files such as models and logs.