fossasia/visdom
Visdom
Visdom is a tool for scientific experiment tracking and real-time data monitoring. It provides a programmatic interface for broadcasting live plots, rich media, and training metrics from scripts to an interactive web dashboard.
The project specializes in high-dimensional data analysis, offering capabilities to project complex feature sets into 2D space using t-SNE and visualize PyTorch model embeddings. It organizes visualizations into named environments, allowing users to isolate different experimental runs and compare plots across these environments in a single view.
The system covers a broad range of dashboard capabilities, including programmatic plot generation for scatter and surface plots, rich media rendering for video and audio, and the creation of editable property panes. It also provides layout state management to serialize window positions and server access controls to restrict network access.
Features
- Programmatic Plotting APIs - Provides a programmatic interface to render scatter, line, bar, histogram, heatmap, and 3D surface plots.
- Experiment Tracking - Organizes and compares different experimental runs by partitioning visualizations into named environments.
- Model Visualization Tools - Visualizes high-dimensional embeddings and training metrics using PyTorch and Numpy to analyze model performance.
- Dimensionality Projection Plots - Projects high-dimensional features into 2D space using t-SNE with interactive lasso selection for analysis.
- Real-Time Plot Updates - Sends live plots, images, and text to a dashboard for real-time monitoring of scientific experiments.
- Environment Namespaces - Provides named logical groups to isolate and organize different experimental runs and their associated visualizations.
- Real-Time Monitoring Dashboards - Broadcasts live plots and rich media from scripts to a web dashboard for monitoring scientific experiments.
- Interactive Dashboards - Builds web interfaces with property panes and event callbacks to manipulate data visualizations in real time.
- Visual Workspace Isolations - Allows users to partition the dashboard into named environments to isolate different experimental runs and share them via unique URLs.
- WebSocket State Synchronization - Maintains a real-time bidirectional connection to push plot updates and event triggers between the server and browser.
- State Serialization - Serializes window positions and plot configurations as portable objects to enable saving, forking, or deleting environment snapshots.
- Server Access Controls - Implements server-side access policies including authentication requirements and restrictions to localhost.
- Server-Side Interaction Callbacks - Implements mechanisms that map browser interaction events to the execution of specific Python callback functions on the server.
- Comparison Views - Overlays plots with identical titles from multiple environments into a single view to analyze differences across experimental runs.
- Property Editing Panels - Creates grids of input fields in the interface to change plot parameters dynamically.
- Event Handling - Triggers server-side callbacks when users perform browser actions like clicking images or updating properties.
- Programmatic Element Construction - Generates dashboard elements and property panes programmatically via a remote API.
- Rich Media Loggers - Displays images, video, audio, and SVG objects within a web dashboard for live monitoring.
