runicorn 0.7.2

Local experiment tracking with model versioning - self-hosted W&B alternative with Artifacts

Runicorn

English | 简体中文

Local-first experiment tracking for teams and individuals who keep ML runs on their own machines or GPU servers.

Runicorn combines a Python SDK, Web UI, Remote Viewer, and optional Windows desktop build into one workflow. It is designed for people who want experiment tracking without pushing runs into a SaaS platform, mirroring remote run folders back to a laptop, or rewriting an existing training stack from scratch.

If your experiments run on remote Linux boxes, Runicorn can start a lightweight Viewer on the remote machine and tunnel it over SSH, so you inspect runs directly where they live.

Documentation Site

---

Why Runicorn

• Data stays under your control: runs, logs, assets, and settings remain on storage you control
• Remote Viewer without file sync: inspect remote GPU-server experiments over SSH instead of copying run folders back first
• Low-friction adoption: start with rn.init(...) and run.log(...), while keeping compatibility with print(), Python logging, torchvision MetricLogger, ImageNet meters, TensorBoard, and tensorboardX
• Real experiment context: store config metadata, dataset references, pretrained references, code snapshots, and archived outputs alongside scalar metrics
• Rich local UI: path tree, compare mode, assets repository, recycle bin, import/export, GPU telemetry, and remote session management

---

Who It's For

Good fit

• You train locally or on your own servers and want experiment tracking without SaaS lock-in
• You often switch between a local workstation and remote GPU machines
• You want metrics, logs, code context, and experiment assets tied to the same run
• You already have training code and want an incremental adoption path instead of a large rewrite

Not a good fit

• You need a hosted collaboration platform, cloud dashboards, or team permissions out of the box
• You only want a minimal CSV logger and do not care about browsing runs later
• Your workflow depends on a managed online ecosystem rather than local or self-controlled infrastructure

---

Quick Start

Install Runicorn:

pip install -U runicorn

Log your first run:

import runicorn as rn

run = rn.init(
    path="cv/resnet50/baseline",
    alias="trial-01",
    capture_console=True,
)

for epoch in range(1, 11):
    train_loss = 1.0 / epoch
    val_acc = 0.70 + epoch * 0.02
    run.log({"train_loss": train_loss, "val_acc": val_acc})

run.summary({"notes": "first stable run"})
run.finish()

Open the viewer:

runicorn viewer

Then visit http://127.0.0.1:23300.

---

What Makes It Different

Workflow	Common friction	Runicorn
Manual local folders + scripts	Metrics, logs, configs, and outputs drift apart	Keeps run history, summary, logs, and assets tied to one run
Remote SSH + tail + ad-hoc plotting	Slow to inspect, hard to compare, easy to lose context	Remote Viewer gives a structured UI over SSH without a sync-first loop
Hosted experiment tracker	Requires internet/service trust and external storage	Keeps experiment data local and under your storage boundary

---

Remote Viewer

Runicorn is strongest when training happens on remote GPU servers.

With Remote Viewer, you can connect over SSH, choose the remote Python environment, start a remote Viewer session, and inspect remote runs through your local browser or desktop app.

No local mirror. No manual sync step. No waiting for large run folders to copy.

runicorn viewer
# Open the Remote page -> connect over SSH -> choose environment -> start session

	Old sync-heavy workflow	Remote Viewer
Wait time	Minutes to hours	Seconds
Local storage copy	Required	Not required
Experiment inspection	Delayed	Direct over SSH

---

Logging Compatibility

Runicorn is built to fit into training code that already exists.

If your loop already uses torchvision-style metric logging, you can often start by replacing one import:

from runicorn.log_compat.torchvision import MetricLogger as MetricLogger

Runicorn also provides compatibility helpers for:

• runicorn.log_compat.imagenet
• runicorn.log_compat.tensorboard
• runicorn.log_compat.tensorboardX

You can also capture print() output and Python logging, but the main value here is structured metric compatibility that turns existing training signals into real experiment curves.

---

Assets and Run Context

Runicorn does not stop at scalar metrics. You can also record the surrounding context that makes runs understandable later:

• config metadata
• dataset references
• pretrained model references
• code snapshots
• archived output files
• cross-run asset browsing and preview

Use run.log_config(...), run.log_dataset(...), and run.log_pretrained(...) inside a run to attach reusable training context directly to the experiment record.

---

Documentation

• Quick Start
• Remote Viewer Guide
• Python SDK Overview
• Web UI Overview
• CLI Overview
• Changelog

---

License

MIT. See LICENSE.