SwanLab

SwanLab is a model training and R&D tool launched by Emotion Machine. It provides training visualization, automatic log recording, hyperparameter recording, experiment comparison and multi-person collaboration capabilities, helping to improve R&D efficiency.

AI Research Tools#Training Visualization#Experiment Management#MLOps Tools

Updated: 2026-10-26

What is SwanLab?

SwanLab is an AI model training and development tool launched by Emotion Machine (Beijing) Technology Co., Ltd., positioned as an advanced AI experiment platform for the AGI era. Targeting the AI training process, it provides professional capabilities such as training visualization, automatic log recording, hyperparameter tracking, experiment comparison, and multi-person collaboration. Developers only need a small amount of code to access it, and it supports more than 30 mainstream training frameworks, making the training process recordable and reviewable.

Who is it for?

Researchers in universities and research institutes need to systematically record experiments and use visualization to find the direction of iteration. Charts are more intuitive than logs.
Enterprise algorithm teams hope to break down communication barriers, synchronize progress with experiment links, and improve collaboration efficiency. Remote collaboration can also be smooth.
Developers using frameworks such as PyTorch and Transformers can quickly start tracking with just six lines of code. The integration cost is negligible.
People who often work remotely or on the go can check the training progress on their phones at any time with multi-device adaptation. They can also keep track of the status during their commute.
For environments without internet access, offline mode can be enabled to meet the recording needs of localized training. Offline scenarios are also traceable.

Key features

Training visualization presents real-time training progress through various charts such as line charts and media charts to assist in analysis and decision-making. The real-time dashboard enables earlier detection of anomalies.

Automatic log collection gathers terminal logs and automatically records hardware information as well as programming environments such as Python and Git. Environmental information is highly valuable for traceability.

Hyperparameter recording can summarize the configuration of each experiment and support search, making it convenient to compare the differences between different experiments horizontally. The parameters are clear at a glance for review.

Experiment comparison presents multiple sets of results in a centralized table, helping researchers draw inspiration and approach better models. It is easier to find directions through comparison.

Multi-person collaboration allows different trainers to run experiments under the same project, and the organization edition can accumulate the team's training experience. Collaboration breaks communication barriers.

How to use SwanLab

Before starting training, use swanlab.init to create a project and write hyperparameters, and record subsequent indicators with swanlab.log.
Don't skip the initialization step.
When running multiple experiments in parallel, use a centralized table for comparison, select the direction first, and then invest more computing power.
Avoid wasting computing power on inefficient experiments.
For team projects, it is recommended to enable the organization edition to consolidate training experience into reusable collaborative assets.
Consolidating experience is more efficient than working alone.
Pay attention to the automatic recording of hardware information.
When training abnormalities occur, you can quickly locate problems at the video memory or environment level.
Fast positioning enables quick loss mitigation.
Enable dark mode when using at night to protect your eyes when staring at the training dashboard for a long time.
The dark interface reduces visual fatigue.