GU143312/dsh-training-lens
A local, read-only training dashboard for dsh that turns existing TensorBoard scalar logs into a focused progress view without touching your trainer.
Pick a primary and a secondary scalar — reward, validation loss, accuracy, learning rate or a custom tag — and state explicitly whether higher or lower is better; by default the panel shows the observed range without declaring improvement, which avoids the usual fake-progress reading of an unlabelled chart. It offers raw values or optional EMA smoothing, per-metric steps, expandable diagnostics, a Chinese/English UI, responsive layout, automatic refresh and run selection, all reading the TensorBoard event files your trainer already writes.
Install
dsh plugin --profile desktop add github:GU143312/dsh-training-lens#v0.1.0The GitHub install line from the README, pinned to v0.1.0 as printed there (the npm package dsh-training-lens 0.1.0 also resolves to this repo). The README shows a release-tarball URL form as well, and a lens-test profile form for a local build.
Compatibility
Requires DSH 0.2.1-rc.2 (tested), Node 22+ and Python 3.10+ with TensorBoard installed. The README is explicit that it works even when the TensorBoard web server is not running — no browser extension, front-end build or CDN is involved. Stated limits: it starts, inspects and stops only its own dashboard, with no trainer controls, telemetry or cloud upload, and compatibility outside the tested combination is not claimed.
Details
- Repo: GU143312/dsh-training-lens
- Category: Learning & Research
- Stars: 0
- Version: 0.1.0; github tag v0.1.0
Alternatives
wangteng2/dsh-learning-english · LUK-jiao/dsh-wait-learn · Flan246/dsh-research-writing