Phase 6: local Streamlit UI for the toolkit

[mrjunos]

Summary

Wraps the existing CLI pipeline in a local Streamlit + Plotly app a non-technical user can run end-to-end: prepare data (tray capture + visual cropping), train models with real-time charts, evaluate, predict on a single bean, and view settings.

Closes #5.

What's in

• ui optional-deps group (streamlit + plotly), almendra ui CLI subcommand, make ui Makefile target.
• Six pages behind a sidebar radio:
  • 🏠 Home — dataset stats, recent runs, health check, inline wizard.
  • 📷 Tray Capture — side-A / optional side-B uploaders, TraySpec form, side-by-side original ↔ rectified+overlay preview, save crops to data/raw/proprietary_tray/sessions/<id>/.
  • 🧠 Train — form (backbone, epochs, lr, image size, pseudo-views), Start/Stop, live Plotly chart of train_loss + val_macro_f1 tailing outputs/<run>/live_metrics.jsonl. Advanced controls (gated fusion, view-dropout, augmentation) behind an expander.
  • 📊 Evaluate — checkpoint picker, metrics table, confusion-matrix heatmap, mis-classified bean gallery.
  • 🚀 Predict — upload one bean photo, run latest ONNX (prefers INT8), show predicted class + confidence + Top-3 + accept/reject verdict.
  • ⚙️ Settings — canonical taxonomy, data sources, current config, paths.
• Bilingual ES/EN from v1 — every visible string lives in ui/components/i18n.py. Sidebar radio toggles between languages. Default: Spanish.
• Live training charts wiring — train.loop now writes one JSONL line per epoch to a metrics file. Controlled by the ALMENDRA_LIVE_METRICS env var, so the CLI use case is untouched.
• Decoupled, file-based UI — stateless across reruns; runs, checkpoints, ONNX files and metrics are discovered from disk under outputs/. Anything that writes the same JSONL schema interoperates.

Decisions locked (from issue #5)

Question	Choice
Stack	Streamlit + Plotly
Language(s)	Bilingual ES/EN toggle from v1 (Spanish default)
Wizard / advanced controls	Inline wizard on Home, Advanced collapsed on Train

Test plan

• streamlit.testing.v1.AppTest smoke-renders all 6 pages in ES and EN (12 cases) without exceptions.
• i18n dictionary check — every key has both languages.
• Unit tests for the metrics-writer / metrics-reader file protocol.
• ruff check + ruff format --check clean.
• 45/45 tests pass in the suites that don't require torch/onnxruntime.
• One end-to-end manual run from the UI (tray photos in → trained model out) — requires real Arabica capture data; deferred to Phase 3.

Notes for review

• Training is launched as a subprocess (not in-thread) so the UI can survive long runs without blocking, and so killing the browser tab doesn't kill training.
• cmd_ui defers the streamlit import until the subcommand is invoked, so the CLI still works without the ui extra installed.
• All UI image rendering uses use_container_width=True (Streamlit's modern replacement for the deprecated use_column_width).

v2 (Phase 6.1) — out of scope here

• Dedicated Labelling page with hotkeys + IAA reporting.
• Export & Bench page.
• Model-package zip exporter (ONNX + INT8 + model card + manifest snapshot).
• "Demo mode" using the public Roboflow data baseline.

🤖 Generated with Claude Code