Day 28 - March 07, 2026

Parent: Week 25 - Daily Log | Nav Order: 28

Summary

Implemented the Legend gallery child app, fixed Gradio file-serving and rendering issues, added docs and tests, and enabled click-to-detail for dataset images.

Files

Added: apps/legend_gallery.py - LegendGalleryApp(BaseApp) (gallery, filters, click-to-details)
Added: docs/legend_app.md - short user/developer guide
Modified: app.py - whitelisted dataset paths for Gradio; registered static paths
Modified: config.yaml - registered legend_gallery under apps

Key Implementation Notes

Data sources: labels from vishal.csv (via settings.storage.get_labels_dir()), images from settings.storage.get_faces_dir() (e.g. {Photo_No}.png).
R_ values (floats) bucketed into levels 0–3; UI supports All level.
Gallery is capped to 30 items (MAX_GALLERY_ITEMS) for responsiveness.
Click behavior: selecting an item shows full image + metadata (photo id, R_ values, demographics) in the side panel.
No absolute paths - all resolution uses platform settings APIs.

Tests Performed

CSV sanity: vishal.csv loads (2,864 rows).
Instantiation test:

.venv\\Scripts\\python.exe -c "from apps.legend_gallery import LegendGalleryApp; inst=LegendGalleryApp('legend_gallery'); print(inst.faces_path, inst.csv_path.exists())"

Functional filter test (programmatic): items, status = inst.load_gallery('R_1','All','All','All') - returns 30 items and a correct status message.
Verified no import/runtime errors after edits.

Issues & Fixes

Gradio InvalidPathError when serving files outside CWD.
- Fix: added launch_kwargs['allowed_paths'] with settings.storage.get_faces_dir() in app.py and called gr.set_static_paths([str(settings.storage.repos_dir)]) in create_main_interface().
Duplicate-block render error (“A block with id … already rendered”).
- Fix: moved detail components into the main layout (side column) and update them via gallery.select() instead of rendering twice.

Behavior Changes

The app can serve images from the dataset directory (whitelisted for Gradio).
Clicking a gallery image displays detailed metadata from vishal.csv.
legend_gallery appears as an app tab (registered in config.yaml).

Commands

Start the platform (development):

uv run gradio app.py

Quick import/test:

.venv\\Scripts\\python.exe -c "from apps.legend_gallery import LegendGalleryApp; inst=LegendGalleryApp('legend_gallery'); print(inst.faces_path, inst.csv_path.exists())"

Recommendations & Next Steps

Make vishal.csv filename and image extension configurable via config.yaml.
Add an assets mode (pre-built assets/legend/R_*/) for Spaces-friendly deployment.
Add unit tests for load_gallery and the click handler.
Implement Level 2: side-by-side comparison view and deeper metadata panels.

Progress: Level 1 architecture complete; model integration (float→level mapping) still in validation.

Detailed Notes

Core Architectural Finalization (Level 1)

We transitioned the project from a conceptual design to a functional, modular Child App within the Gradio-based ecosystem. Key points: - Dynamic Module Integration: Added `LegendGalleryApp` as a `BaseApp` subclass so the main `app.py` discovers and renders it as a first-class tab. - Optimized Data Engine: Implemented a lazy-loading cache for `vishal.csv` (`_get_df`) to avoid redundant disk I/O across UI interactions. - Normalized Severity Mapping: Introduced `LEVEL_RANGES` to bucket continuous R\_ floats into Levels 0–3 for clinical-readability.

Strategic "Loophole" Remediation

We identified and patched several stability and environment gaps: - Memory & Performance: Enforced `MAX_GALLERY_ITEMS = 30` to prevent browser overload when rendering many high-resolution faces. - Hot-reload stability: Hardened `discover_apps` in `app.py` to handle class reloads during development safely. - Environment parity: Detect whether running on Hugging Face Spaces vs local dev to adjust static paths and disable hot-reloads where appropriate.

Data & Labeling Alignment

- Region descriptions: Added human-readable `REGION_DESCRIPTIONS` for R_1..R_9 so clinicians see context when selecting regions. - Demographic filters: Filters for `I_1` (Gender) and `I_2` (Age Groups) are enabled and tested against CSV metadata. - Pathing: All file resolution uses `settings.storage` APIs—no hard-coded absolute paths.

Current Project Status

Milestone	Status	Notes
Level 1 Architecture	✅ COMPLETE	Class-based gallery, filtering, click-to-details implemented
Data Sync Logic	✅ STABLE	CSV caching and face path resolution via `HuggingFaceManager`
Model Integration	🛠️ IN PROGRESS	Validating float→level mapping and thresholds
Level 2 Design	📋 PLANNED	Side-by-side comparison & metadata deep-dives

Next Steps for Tomorrow

Run a Level 1 smoke test across the full dataset to validate image serving and filter combinations.
Add configuration options for vishal.csv filename and image extension in config.yaml.
Begin Level 2 design: implement side-by-side comparison view and deeper metadata panels.
Add unit tests for load_gallery and the click handler; automate in CI if possible.

The Legend gallery is integrated and stable for Level 1 usage; the next focus is on formalizing model thresholds and Level 2 UX features.