UI Pilot Dashboard Guide¶

v0.2.0 | Gradio-based Pilot Dashboard for lightweight inference and feedback

Overview¶

The Pilot Dashboard is a streamlined Gradio interface focused on rapid inference and human feedback collection. It's lighter than Studio and ideal for: - Quick model evaluation - Human annotation workflows - Correction routing to the continual learning pipeline

Launch: Included in adaptshot-studio or as standalone component

Screenshot 7: Pilot Dashboard Full View¶

Dimensions: 1280×900 What to capture: The complete Pilot Dashboard with all panels visible

UI Layout¶

┌─────────────────────────────────────────────────────────┐
│                   AdaptShot Pilot v0.2.0                  │
├──────────────────────┬──────────────────────────────────┤
│                      │                                  │
│   Query Image        │   Prediction Result              │
│   ┌──────────────┐   │   ┌──────────────────────────┐   │
│   │              │   │   │ Class: cat               │   │
│   │  [Image]     │   │   │ Confidence: 87%          │   │
│   │              │   │   │ ACT: ACCEPTED            │   │
│   │              │   │   │ OOD: No                  │   │
│   └──────────────┘   │   └──────────────────────────┘   │
│                      │                                  │
│   [Upload] [Camera]  │   Conformal Set                  │
│                      │   ┌──────────────────────────┐   │
│                      │   │ [cat] [dog]               │   │
│                      │   └──────────────────────────┘   │
│                      │                                  │
│                      │   Uncertainty                    │
│                      │   Epistemic:  ████░░ 0.08       │
│                      │   Aleatoric:  ██████ 0.12      │
│                      │   Composite:  █████░ 0.12      │
├──────────────────────┴──────────────────────────────────┤
│   Support Info: 20 images | 3 classes | Buffer: 85/200   │
├─────────────────────────────────────────────────────────┤
│   [⚙️ Config]  [📊 Diagnostics]  [💾 Save]  [📦 Export] │
└─────────────────────────────────────────────────────────┘

Component IDs (for Gradio event wiring)¶

Component	`gr` ID	Event
Query image upload	`query_input`	`.change()` → triggers prediction
Camera capture	`camera_input`	`.change()` → triggers prediction
Prediction label	`pred_label`	Output from `predict_fn`
Confidence gauge	`conf_gauge`	`.plot()` update
ACT badge	`act_badge`	Color-coded label
Conformal set	`conformal_tags`	`.update(value=tags)`
Uncertainty bars	`uncertainty_plot`	`.plot()` with 3 bars
Support info bar	`status_bar`	`.update(value=status_text)`
Save button	`save_btn`	`.click()` → `save_fn`

Screenshot 8: Correction Routing Workflow¶

Dimensions: 1280×900 (sequential; capture 3 states side by side)

State 1: Prediction Requires Feedback¶

What to show: A prediction where ACT returns REQUEST_FEEDBACK (orange badge)

Query image visible on left
Prediction shown but with orange "⚠ NEEDS REVIEW" badge
Confidence at 62% (below threshold)
Correction panel visible on right side

State 2: Correction Input¶

What to show: The correction form being filled

Field	Content
Predicted label	"dog" (disabled, gray)
True label dropdown	"cat" (user selected)
Confidence slider	0.9
Submit button	Highlighted blue

State 3: Correction Confirmed¶

What to show: Post-correction state

Green banner: "✅ Correction submitted successfully"
Buffer size updated (e.g., "86/200")
Calibration status shows "Updated"
Model ready for next prediction

Event Wiring Diagram¶

graph TB
    A[query_input.change] --> B[predict_fn]
    C[camera_input.change] --> B
    B --> D[pred_label]
    B --> E[conf_gauge]
    B --> F[act_badge]
    B --> G[conformal_tags]
    B --> H[uncertainty_plot]

    I[correct_btn.click] --> J[correct_fn]
    J --> K[status_bar]
    J --> L[correction_status]

    M[save_btn.click] --> N[save_fn]
    N --> K

    O[config_btn.click] --> P[open_config_modal]

Configuration Panel (accessible via ⚙️ button)¶

Field	Type	Default
Backbone	Dropdown	`resnet18`
Inference Mode	Radio	`prototypical`
Conformal Alpha	Slider (0.01-0.20)	`0.05`
OOD Detection	Toggle	ON
Explainability	Toggle	ON
Buffer Size	Number input	`200`
Temperature	Number input	`1.0`

Usage Examples¶

Quick Evaluation Workflow¶

Launch Pilot Dashboard
Click "📁 Load Support Set" → select folder
Upload query image or use camera
View prediction, confidence, uncertainty
Click "✏️ Correct" if prediction is wrong
Repeat for additional queries

Annotation Workflow¶

Load support set
For each unlabeled image: a. Upload image b. Review prediction c. Confirm or correct the label d. Correction is automatically routed to the learner
Periodically save checkpoints

Troubleshooting¶

Issue	Solution
Camera not working	Grant browser camera permissions; use file upload as fallback
Slow predictions	Switch to `mobilenet_v3_small` backbone
UI elements overlapping	Resize browser window to ≥1280px width
Correction not saving	Check disk permissions for checkpoint directory