marcellorusso/orchid-ncd-models

OrchID-NCD Models

Trained model weights for the OrchID-NCD project — ultra-fine-grained visual classification of Ophrys orchids.

Fine-grained classification of six cryptic Ophrys species (O. exaltata, O. garganica, O. incubacea, O. majellensis, O. sphegodes, O. sphegodes Palena) using ResNet-18/50, ConvNeXt-Tiny/Small, and DINOv2-Small/Base.

Exp 6: Cross-Entropy

5-fold stratified cross-validation on deduplicated clean split (2,232 train, 300 test).

Training recipe: epochs=100, patience=15, effective batch=32 (gradient accumulation ×4), per-architecture LR from registry.

Model	Val F1 (macro)	Test F1	Test Acc	Best Fold	Folds
ResNet-18	0.9006 ± 0.0088	0.8799	0.8800	2	5
ResNet-50	0.9178 ± 0.0062	0.8822	0.8800	3	5
ConvNeXt-Tiny	0.8918 ± 0.0112	0.8902	0.8900	1	5
ConvNeXt-Small	0.8828 ± 0.0140	0.8626	0.8633	4	5
DINOv2-Small	0.9061 ± 0.0161	0.8883	0.8867	1	5
DINOv2-Base	0.9157 ± 0.0079	0.9271	0.9267	1	5

Exp 7: Supervised Contrastive Learning

Two-phase training: SupCon pretraining (InfoNCE, τ=0.07) → CE fine-tuning with frozen backbone.

Recipe: same per-architecture optimizer/LR, projection dim=128, CE Phase LR=0.01, patience=15 on val metrics.

Model	Val F1 (macro)	Test F1	Test Acc	Best Fold	Folds
ResNet-18	0.6123 ± 0.0744	0.7357	0.7367	4	5
ResNet-50	0.7359 ± 0.0640	0.7993	0.7967	1	5
ConvNeXt-Tiny	0.8755 ± 0.0191	0.6810	0.7133	1	5
ConvNeXt-Small	0.8821 ± 0.0103	0.6886	0.7133	1	5
DINOv2-Small	0.9058 ± 0.0143	0.8958	0.8967	4	5
DINOv2-Base	0.9135 ± 0.0061	0.6874	0.7200	3	5

Exp 8: Spherical Orthogonal Prototypes (SpHOR)

Two-phase training: SupCon + Spherical Orthogonal Prototypes → CE fine-tuning with frozen backbone.

Recipe: same as Exp 7, with spherical prototype repulsion (repulse=0.01).

Model	Val F1 (macro)	Test F1	Test Acc	Best Fold	Folds
ResNet-18	0.5563 ± 0.1110	0.5369	0.5800	4	5
ResNet-50	0.7361 ± 0.0772	0.6093	0.6600	1	5
ConvNeXt-Tiny	0.8755 ± 0.0191	0.6810	0.7133	1	5
DINOv2-Small	0.9059 ± 0.0066	0.7470	0.7467	4	5

Structure

exp6_ce/                           — Exp 6 (Cross-Entropy)
  exp6_clean_split_resnet18/
    results.json                   — aggregated metrics + test results
    config.json                    — training configuration
    best_fold_N.pt                 — weights of the best fold
    best_fold_N_adapter/           — LoRA adapter (DINOv2 only)
exp7_supcon/                       — Exp 7 (SupCon + CE fine-tune)
  exp7_supcon_resnet18/
    ...
exp8_sphor/                        — Exp 8 (SpHOR)
  exp8_sphor_resnet18/
    ...

Usage

The classifier in the OrchID-NCD Space downloads these weights at startup and uses them for inference.

Links

• Dataset
• Live Demo
• GitHub