Commit History
-
feat: vis heatmap in --vis (no new arg)
Verified signatureAdd get_vit_heatmap() per-patch logits 21x21->336 jet overlay where ViT looks (MoE 5 prefix tokens). --vis now 1x3: image | heatmap overlay | bar (was 1x2). Uses core.head on patch tokens, top class mean. Tested test.png 1.2MB with heatmap, noise 62K. Autocast bf16, fallback to 1x2 if heatmap fails. No new CLI arg, enhanced existing --vis.253e323 -
fix: gate unknown on conf to avoid dummy-centroid false UNKNOWN on high-conf Eucalyptus
Verified signatureDummy 999x512 centroids give sim -0.02 for real Eucalyptus cornuta 98% (placeholder built from muted green). Previous rule flagged very far even at 98% conf. Fix: require sim<thr AND conf<75% for UNKNOWN, high conf overrides dummy distance. Added note '(high conf overrides low sim - dummy centroids, rebuild for true)'. Verified: test.png Eucalyptus 98% now KNOWN, noise 42% UNKNOWN, green 96% KNOWN.b799883 -
feat: centroid-based unknown species detection (2MB cache)
Verified signatureAdd 999x512 centroid cache (~2MB raw, 1MB compressed) for open-set detection. New: CENTROIDS_PATH, load_centroids(), centroid_similarity(), detect_unknown() ensemble (cosine + MSP + margin + energy). predict_images now returns is_unknown + unknown_info, identify/batch show ⚠️ warning + sim/conf/margin. New cmds: build-centroids (--dummy placeholder 932KB, --manifest for true per-class mean) and test-unknown. Tested RTX4050: known green sim 0.766 conf 96% KNOWN, noise 0.437 UNKNOWN, checker 0.450 UNKNOWN, text 0.448 UNKNOWN. Batch CSV now includes is_unknown,sim_max.54a4bbb -
fix: CLI public release fallback to species_labels.json + safetensors
Verified signaturePublic HF repo missing private data/wa_plants_200k/manifest CSVs and pt ckpts. Patch load_species_map to use species_labels.json, add _resolve_ckpt() for PlantDetect-*.safetensors, safetensors.torch.load_file support, Dense/MoE auto-detect, robust cmd_info. Tested: info + single/multi-view on RTX4050 BF16/FP8 AdaRound 104MB.cdc1714
-
Add standalone plant_cli.py for field and batch plant identification
Verified signatured499940