Next-gen ML experiment tracking
The ultimate trailer for your deep learning experiments — a lightning-fast, open-source experiment tracker combining TensorBoard’s depth, Aim’s fluidity, and W&B’s ecosystem, powered by a high-performance Rust core and a local-first philosophy.
Why Trailer?
Section titled “Why Trailer?”High-performance Rust core
LTTB downsampling, batched ingestion, 100k-entry ring buffer — P99 < 100µs.
All data types
Scalars, text, figures, tables, media, histograms, embeddings, PCA, model graphs, and loss landscapes.
Loss landscape visualization
Filter-normalized 2D loss surfaces — heatmap, contours, and an interactive 3D surface with gradient-descent playback.
Three storage backends
SQLite (default), PostgreSQL (teams), and TensorBoard-style file mode.
One-command local mode
`trailer up` starts the full dashboard — no config, no HTTP overhead.
Cross-project Explore
Line, scatter, and parallel-coordinates charts across runs — saved and shareable.
Multi-theme UI
6 preset themes plus a theme-builder for custom accent colors.
Quick start
Section titled “Quick start”pip install trailer-sdkfrom trailer import Tracker
t = Tracker(project="demo")for step in range(10): t.log({"loss": 1.0 / (step + 1)}, step=step)t.finish()trailer up # open http://127.0.0.1:5120How it works
Section titled “How it works”Data is buffered in a ring buffer and flushed in batches through PyO3 (local) or HTTP (remote) into the Rust core, then served to the web UI.