AdaptShot Studio: Complete Guide

v0.2.0 | Comprehensive guide to the Gradio-based Studio Dashboard

---

Overview

AdaptShot Studio is a Gradio web interface for interactive few-shot learning. It provides visual configuration, support set management, inference with diagnostics, and human feedback integration — all without writing code.

Launch: adaptshot-studio from the command line, or python -m adaptshot.studio.app

---

Screenshot 1: Dashboard Hero

Dimensions: 1920×1080 (full viewport) What to capture: Top section of the Studio interface showing the main configuration panel

UI Elements (top to bottom)

Element	Position	Description
Title Bar	Top center	"AdaptShot Studio v0.2.0" with version badge
Backbone Selector	Upper-left card	Dropdown: resnet18 or mobilenet_v3_small
Inference Mode	Next to backbone	Radio buttons: Nearest Neighbor / Prototypical / Contrastive
Eco Mode Toggle	Top-right card	Toggle switch with leaf icon; shows estimated energy savings
Similarity Metric	Below backbone	Dropdown: cosine or euclidean
Device Indicator	Bottom of config card	Shows CPU with green dot when active
Load Support Set Button	Center action bar	Primary button: "📁 Load Support Set"
Status Bar	Bottom	"Ready" / "Initialized with N classes"

Expected Data

• Backbone: resnet18 (default)
• Inference mode: Prototypical selected
• Eco mode: OFF
• Metric: euclidean

Color Reference

• Primary: #1a73e8 (Google Blue)
• Background: #f8f9fa
• Cards: #ffffff with subtle shadow
• Status green: #34a853

---

Screenshot 2: Support Set Loading

Dimensions: 1920×1080 (full viewport) What to capture: The file upload interface with loaded support set and class distribution

UI Elements

Element	Position	Description
File Upload Area	Left panel (60%)	Drag-and-drop zone with "+" icon and "Drop images here or click to browse"
Folder Path Input	Below upload area	Text input for directory path; "Scan Folder" button
Label Strategy Selector	Below folder input	Dropdown: "From folder names" / "From filenames" / "Manual assignment"
Image Grid	Center panel	Thumbnail grid of loaded support images (4×N grid, 120px thumbnails)
Class Distribution	Right panel (40%)	Bar chart showing images per class
Class Labels	Below chart	List: "cat: 10 images", "dog: 10 images", etc.
Confirm Button	Bottom right	"✅ Confirm & Initialize" (blue, 200px wide)
Cancel Button	Next to confirm	"Cancel" (gray outline)

Expected Data

• 20 images loaded across 3 classes
• Class distribution: cat=8, dog=7, bird=5
• Folder path: /data/support_set/

---

Screenshot 3: Inference Results

Dimensions: 1920×1080 (full viewport) What to capture: Query image with prediction results, confidence gauges, and nearest neighbors

UI Elements

Element	Position	Description
Query Image	Left panel (50%)	Uploaded query image displayed at 400×400
Prediction Label	Above image	Bold text: "Prediction: cat" with confidence percentage
Confidence Gauge	Right of prediction	Semi-circular gauge (0-100%) with color gradient (red→yellow→green)
Calibration Info	Below gauge	"Raw: 0.92
ACT Decision Badge	Below calibration	Green badge "✓ ACCEPTED" or orange "⚠ NEEDS REVIEW"
Conformal Set	Right panel (50%)	Card: "Conformal Set (95% coverage)" with class tags
Uncertainty Panel	Below conformal	Three horizontal bars: Epistemic (blue), Aleatoric (orange), Distributional (red)
Nearest Neighbors	Bottom panel	Horizontal scroll of 5 support image thumbnails with similarity scores
Explain Button	Bottom right	"🔍 Explain Prediction" link
Correct Button	Next to explain	"✏️ Correct" opens feedback modal

Expected Data

• Prediction: dog, confidence 87%
• ACT: ACCEPTED
• Conformal set: [dog, cat] (2 classes)
• Uncertainty: epistemic=0.08, aleatoric=0.12, distributional=0.05
• Nearest neighbor: support example #3, similarity=0.92

---

Screenshot 4: Human Feedback Panel

Dimensions: 800×600 (modal overlay) What to capture: The correction input modal with feedback options

UI Elements

Element	Position	Description
Modal Title	Top	"Submit Correction" with ✏️ icon
Current Prediction	Below title	"Model predicted: dog (0.87)" in gray text
True Label Input	Center	Text input or dropdown of known classes
Confidence Slider	Below label	Range 0.0-1.0 with "How confident are you?" label
Slider Value	Right of slider	Shows current value (e.g., 0.90)
Comparative Option	Below slider	Checkbox: "This was a comparative judgment"
Alternative Label	Conditional	Appears when comparative is checked; dropdown of other classes
Submit Button	Bottom right	"Submit Correction" (blue)
Cancel Button	Next to submit	"Cancel" (gray)
Buffer Status	Bottom left	"Buffer: 102/200" with thin progress bar

Expected Data

• Current prediction: dog
• True label input shows "cat" being typed
• Confidence slider at 0.9
• Buffer: 102/200 (51%)

---

Screenshot 5: Diagnostics Dashboard

Dimensions: 1920×1080 (full viewport) What to capture: The monitoring/diagnostics tab with charts and metrics

UI Elements

Element	Position	Description
ECE Chart	Top-left card	Line chart: ECE over time (last 100 predictions)
Current ECE	Above chart	"Debiased ECE: 0.042" with trend arrow
Buffer Memory Gauge	Top-right card	Circular gauge: "Buffer: 85/200 (42%)"
Latency Profile	Middle-left	Bar chart: avg inference time per step (extract, search, calibrate, act)
Calibration History	Middle-right	Scatter plot: raw confidence vs. accuracy
Conformal Coverage	Bottom-left	Line chart: observed coverage vs. target (dashed line at 95%)
OOD Rate	Bottom-right	"OOD Rate: 2.3%" with sparkline
Per-Class Accuracy	Bottom table	Table: class name, support count, accuracy, threshold
Refresh Button	Top right corner	"🔄 Refresh" (updates all charts)

Expected Data

• ECE: 0.042 trending down
• Buffer: 85/200 (42%)
• Latency: extract=45ms, search=8ms, calibrate=2ms, act=1ms, total=56ms
• Conformal coverage: 94.7% (close to 95% target)
• OOD rate: 2.3%
• 12 classes in accuracy table

---

Screenshot 6: Export & Deployment Bundle

Dimensions: 800×600 (modal or panel) What to capture: The export configuration interface

UI Elements

Element	Position	Description
Export Title	Top	"Export Deployment Bundle"
Checkpoint Path	Text field	Pre-filled with last save path
Include Model Head	Checkbox	"Include CA-EWC model head (.pt)" checked
Include Embeddings	Checkbox	"Include support embeddings (.npy)" checked
Include ONNX Model	Checkbox	"Include ONNX backbone (.onnx)" unchecked
Export Format	Radio	"Full Bundle" / "Embeddings Only" / "Config Only"
Estimated Size	Text	"Estimated: 24.3 MB"
Export Button	Bottom right	"📦 Export Bundle"
Download Progress	Center	Progress bar during export (0% → 100%)
Success Message	Green banner	"✅ Bundle exported to adaptshot_bundle_2025.tar.gz"
Download Link	Below message	"Click to download" link

Expected Data

• Checkpoint: ./checkpoints/production_20250115_142300.json
• Model head: 0.5 MB
• Embeddings: 23.8 MB
• Total: 24.3 MB

---

State Transitions

graph LR
    A[Start: No Support] --> B[Support Loaded]
    B --> C[Predicting]
    C --> D{ACT Decision}
    D -->|ACCEPT| E[Result Displayed]
    D -->|REQUEST_FEEDBACK| F[Feedback Modal]
    F --> G[Correction Submitted]
    G --> B
    E --> H[Export / Save]

---

Keyboard Shortcuts

Key	Action
Ctrl+O	Open support set
Ctrl+P	Predict on uploaded image
Ctrl+E	Explain current prediction
Ctrl+S	Save checkpoint
Ctrl+R	Refresh diagnostics
Esc	Close modal / cancel

---

Troubleshooting

Issue	Solution
Studio won't start	Install GUI extras: pip install "adaptshot[gui]"
Images appear black	Ensure images are RGB, not CMYK or grayscale
Predictions are slow	Enable Eco Mode or switch to MobileNetV3 backbone
Charts not updating	Click "Refresh" or reduce max_buffer_size
Export fails	Check disk space; embeddings can be large