Track Experiments with Jupyter Notebook

Combine SwanLab with Jupyter to get interactive visualizations without leaving the Notebook.

Install SwanLab in Notebook

!pip install swanlab -qqq

ps: -qqq is used to control the amount of output information during command execution and is optional.

Logging into Swanlab in Notebook

jupyter --execute is the command-line execution mode provided by Jupyter, commonly used in automation, CI/CD pipelines, and scheduled tasks. In these scenarios, unattended operation is required, and interactive input (such as input() or manually pasting keys) is not allowed; otherwise, the process will stall or encounter errors.

Environment variables automatically inject the API Key into the Notebook execution environment, ensuring that the Notebook runs smoothly in any non-interactive environment. Therefore, you need to log in via swanlab.login.

import swanlab
import os

SWANLAB_API_KEY = os.getenv("SWANLAB_API_KEY")
swanlab.login(api_key=SWANLAB_API_KEY)

INFO

On Kaggle, you should use Kaggle Secrets to store API keys. This prevents key leaks, simplifies management, conforms to security best practices, and ensures that Notebooks can be securely shared and collaborated on.

from kaggle_secrets import UserSecretsClient
user_secrets = UserSecretsClient()
SWANLAB_API_KEY = user_secrets.get_secret("SWANLAB_API_KEY")

import swanlab
swanlab.login(api_key=SWANLAB_API_KEY)

Interact with SwanLab in Notebook

import swanlab

swanlab.init()
...
# In the Notebook, you need to explicitly close the experiment
swanlab.finish()

When initializing the experiment with swanlab.init, a "Display SwanLab Dashboard" button will appear at the end of the printed information:

Clicking this button will embed the SwanLab web page for the experiment in the Notebook:

Now, you can directly see the training process and interact with it in this embedded web page.