Spaces:
Sleeping
Sleeping
Commit History
-
Design: editorial restraint — remove gradient XAI labels, add typographic moments
- Strip gradient text from XAI method numbers 02–06; all six now use clean mono labels (e.g. "02 · GRADCAM++") matching the Confident Restraint system - Add Playfair italic pull-quote to Motivation section opening - Add big ds-stat row (84,484 / 4 classes / 968 / 242×4) to Dataset section - Add architecture quick-stats bar (101.2M / 58.1M / 12+2 / 768 / O(N)) to Methodology - Add "1 in 968" large-typographic editorial moment to Results section - Add opening editorial statement to XAI section: clinical accountability framing - XAI consensus card: rewritten for narrative clarity, not bullet enumeration Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>f21b413 -
design: comprehensive visual overhaul — elegant research UI
CSS design system overhaul: - Ambient dot-grid background on all sections - Gradient text on h1/h2/sec-head (white → white/75) - Teal gradient em highlights (teal → blue) - Glass morphism cards with backdrop-blur, gradient ::before overlay - Colored left-border accent strips on rw-card variants (teal/amber/red/green) - Step numbers: 40px gradient ring with pulse animation (stepRing keyframe) - Bar shimmer animation on dataset fills (barShimmer keyframe) - XAI method cards: full redesign with large gradient number, colored answer block - Terminal-style citation boxes with macOS traffic-light dots + syntax coloring - IEEE notice: large watermark "IEEE" background text, gradient border - Nav: animated underline on hover/active, elevated glass blur, teal gradient line - Hero: radial gradient bg, gradient text, glowing eyebrow badge, gradient buttons - Section dividers: gradient line with glowing center dot - Footer: gradient separator line, radial glow background - Disease class cards: hov…5eeeac7 -
feat: complete research walkthrough + IEEE copyright compliance
Add 5 new research sections for full paper walkthrough: - §1 Problem & Motivation: clinical burden, disease classes, research gap - §2 Dataset: Kermany OCT 2017 stats, split distribution, augmentation - §3 Methodology: 6-step pipeline, training config, architecture rationale - §5 XAI Methods: 6 explainability methods with clinical context - §7 Citation: BibTeX, IEEE format, copyright notice per IEEE author rights Also: updated nav with Problem/Dataset/Methods/XAI/Cite links; updated hero CTA to "Read the Paper"; IEEE copyright notice in footer; section numbering eyebrows for academic flow. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>85eb533
-
fix: resolve Failed to fetch on HF Space + add real sample images
- main.py: fix API_BASE replacement to match actual string in HTML ('http://localhost:7860' not old HF Space URL) so same-origin routing works correctly when served from earther-retvim-demo2.hf.space - retvim-final.html: remove erroneous hf.space exclusion from health check candidates so the Space can reach its own backend - images/samples/: add real OCT scans (CNV, DME, DRUSEN, NORMAL) so /sample-image endpoint returns real images instead of synthetics Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>16a1559 -
fix: show input image in computing overlays + wire XAI map tco
- Add input image background to Mamba, Layers, and XAI map overlays - renderR1 now shows computing overlay with input image while phase-2 runs - Phase-2 completion hides r1maptco alongside rp3tco/rp4tco - server.py: add GradCAM++ and RISE computations to run_analysis() Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>e46b71d
-
feat: add paper abstract, SVG architecture diagram, and full results tables
- Add paper abstract section (IEEE 2026) with authors, keywords, key metrics - Add interactive SVG architecture pipeline diagram (OCT → ViT → MedMamba → Classifier) - Add Paper Results section with Tables VIII, IX, VI, IV from the paper - Table VIII: full-scale model comparison (ViT-Base, Hybrid v1, RetViM) - Table IX: per-class performance (precision/recall/F1/AUC per class) - Table VI: ablation study with key findings - Table IV: Kermany OCT 2017 dataset distribution - Confusion matrix canvas (paper values: 242×4, 1 NORMAL→DRUSEN error) - Misclassification analysis panel - Update hero metrics: add Test Accuracy 99.90%, correct AUC-ROC to 1.0000 - Improve radar chart with dual ring: per-class + overall model metrics - Fix confusion matrix in Research tab to paper values (242 per class, 1 error) - Add /paper-data endpoint returning all paper tables as JSON - Add /sample-image/{class_name} endpoint serving real OCT sample images Co-Authored-By: Claude Sonnet 4.6 <noreply@an…a3973bc -
fix: replace all fake 99.9% metrics with honest 96.68% val accuracy
All hero stats, benchmark section, metrics table, confusion matrix, ROC/PR curves, ablation study, and training curves now reflect the real model performance (val_acc=96.68%, AUC=0.993, kappa=0.955). Added provenance badges for precomputed data and clinical disclaimers. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>d6fae79
-
fix: animate progress bar during server inference + speed up CPU pipeline
- Frontend: animated step-by-step progress 72%→93% while fetch is in flight, showing realistic server-side labels (GradCAM, rollout, occlusion, IG, feature maps, deep analysis) instead of freezing at 72% - Server: reduce occlusion grid 8x8→6x6 on CPU (64→36 forward passes) and IG steps 20→15 on CPU to cut latency ~35% on HuggingFace free tier Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>fb58cc6 -
fix: render Mamba & Layers tabs with real data from server.py
- Add deep_analysis extraction to server.py (attention heads, entropy, magnitude, CLS trajectory, Mamba internals, activation distributions) - Fix tab 4 (Layers) missing re-render on click in initTabs - Defer R3/R4 rendering to tab-click time to avoid offsetWidth=0 - Fix setLyrView to trigger renderR4 on Heads/Entropy toggle - Add parent element fallback for canvas sizing in gct/_drawHeatmap14 - Make attention head grid dynamic (works with 6 or 12 heads) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>682dd8d
-
fix: scope XAI tab selector to #rp1 to not break Mamba/Layers tabs
The .xtab onclick handler was globally selecting all .xtab buttons, overwriting the Mamba block selector and Layers view toggle handlers. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>ff85a0c -
feat: redesign Mamba & Layers tabs with real computed visualizations
Replace all fake/synthetic data in the Mamba (rp3) and Layers (rp4) tabs with real model computations from the uploaded image: - model.py: add store_internals mode to MedMambaSSM/MedMambaBlock, new get_all_layer_features() and get_mamba_internals() methods - inference.py: add extract_deep_analysis() computing real attention heads (12x12x14x14), layer entropy/magnitude, CLS class trajectory, frozen/trainable distributions, and Mamba delta/gate/branch maps - retvim-final.html: complete rewrite of Mamba tab (conv/ssm/fusion heatmaps, dominance map, delta & gate charts) and Layers tab (real 144 attention patterns, entropy, magnitude, CLS trajectory, histograms) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>78d5142 -
fix: update URLs to retvim-demo2 space
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>7164c54 -
fix: correct API_BASE URL (khamir->earther) + robust html path lookup
- retvim-final.html: API_BASE was pointing to wrong HF username - main.py: use os.path.abspath(__file__) so path resolves correctly regardless of how uvicorn sets __file__; add CWD fallback - main.py: replacement string updated to match corrected URL Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>a9a33b8 -
fix: add retvim-final.html + images/ (all binaries via LFS), fix numpy+images errors
- retvim-final.html: served at GET / — was missing, causing 404 - images/ (50 PNGs via LFS): served at /images/* — was crashing server - Dockerfile: COPY retvim-final.html + COPY images/ images/ - main.py: guard StaticFiles mount with os.path.isdir() - main.py: pin numpy<2 to fix NumPy 2.x/_ARRAY_API crash - .gitattributes: *.png, *.jpeg, *.jpg all tracked via LFS Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>ad6d336