See what your model is actually doing — training logs become a plain-English story with a letter grade, live in your editor.

Epochix dashboard: an animated network, a letter grade and a plain-English story

A live dashboard for deep-learning training runs. Drop in a training log (YOLO, PyTorch Lightning, Keras, HuggingFace, FastAI, …) and watch the metrics, the architecture, the phase journey, and a plain-English narrative all update epoch-by-epoch — without ever leaving the editor.

New here? Install, click the Epochix icon in the sidebar, and hit ▶ Try a Demo Run — the dashboard above opens on a bundled training run. No Python, no data, no setup.

What you get

Panel	What it shows
Network State	Auto-detected model architecture (real layer names + param counts), animated activation flow, accuracy ring on the output, overfitting halo
Learning Curve	Per-epoch primary-metric trajectory with grade-threshold reference lines and a smoothing slider
Phase Journey	Awakening → Learning → Understanding → Mastering → Polishing ribbon with per-phase grades
Skill Dimensions	Task-aware radar — mAP50 / Precision / Recall / Localisation for detection runs, Accuracy / Fitting / Generalisation for classification, 1−EER / TAR for biometrics, and so on
Live Metrics	TensorBoard-style scalar cards — latest value, ▲/▼ delta vs previous epoch, gradient-area sparkline of the full trend
Distributions	Parameter share by layer, min/IQR/median/max metric box-summary, value histograms
Engineer Panel	Loss / Accuracy / Overfitting-gap / LR-schedule (log-y) charts with EMA smoothing, log-scale, x:epoch/step toggles, best-epoch markers, and a stats table
Training Diagnostics	Health score gauge + interpreted cards for overfitting, convergence (least-squares slope), stability (σ of Δ), generalisation
In Plain English	Non-technical narrative for the manager / collaborator who just wants to know "is it good?"

Two ways to run

Sidecar mode (recommended — full features)

Install the Python sidecar once:

pip install epochix

Then in VS Code:

Press Ctrl+Alt+M (or Cmd+Alt+M on macOS) to open the dashboard
Right-click any .log file → Epochix: Open Log File…
Or run Epochix: Watch Active Terminal to stream a live training session

The extension spawns the local server, points the webview at it, and streams metrics + architecture into the dashboard live.

Standalone mode (no Python required)

If pip install epochix isn't available, the extension still parses logs in-process using a built-in TypeScript engine. Open a log file exactly the same way; you'll get the core story (grade, phase, narrative, learning curve). Some panels (per-metric series, detected architecture) stay sparse — that data only exists in sidecar mode.

Commands

Command	Default keybinding
`Epochix: Open Dashboard`	`Ctrl+Alt+M` / `Cmd+Alt+M`
`Epochix: Open Log File…`	(right-click any `.log`)
`Epochix: Watch Active Terminal`	—
`Epochix: Export Current Run`	—
`Epochix: Compare Two Runs`	—

Settings

Setting	Default	What it does
`epochix.autoWatchTerminal`	`true`	Auto-open the dashboard when ML training is detected in an integrated terminal
`epochix.taskHint`	`auto`	Force a task type: `classification` / `detection` / `regression` / `biometric` / `gaze` / `nlp`
`epochix.useSidecar`	`auto`	Whether to use the Python sidecar when present (`auto` / `always` / `never`)
`epochix.sidecarPath`	(empty)	Override the path to the `epochix` executable
`epochix.theme`	`auto`	Dashboard theme — `auto` follows VS Code's colour theme
`epochix.locale`	`en`	UI language — `en` / `fa` (RTL) / `fr`

Supported log formats

PyTorch Lightning · Keras / TensorFlow · HuggingFace Trainer · Ultralytics YOLO · FastAI · Accelerate · plus a universal key=value / key: value / JSON fragment fallback for arbitrary logs.

Task auto-detection looks at which canonical metrics show up (mAP for detection, perplexity for NLP, EER for biometric, MAE for gaze, accuracy for classification, etc.) and grades against task-appropriate thresholds.