OrchID NCD — Model Zoo
Trained models for Ophrys orchid classification (6 species), part of the National Center for Biodiversity (NCD) research.
All models trained on the orchid-ncd-dataset (exp6 clean split, zero data leakage).
Serie 1 — Cross-Entropy Loss
Fair comparison of 6 architectures with standard cross-entropy loss. All models share identical training configuration for comparability.
Training Configuration
| Parameter | Value |
|---|---|
| Epochs (max) | 100 |
| Early stopping | Patience=15 on val_loss |
| Batch size | 8 (effective 32 with gradient accumulation ×4) |
| Optimizer | Adam, lr=0.001 |
| Scheduler | ReduceLROnPlateau |
| Validation | 5-fold StratifiedKFold (seed=57) |
| Pretrained | ImageNet |
Models
| Model | Parameters | Type | Status |
|---|---|---|---|
| ResNet18 | 11.7M | CNN | Complete |
| ResNet50 | 25.6M | CNN | In progress |
| ConvNeXt-Tiny | 28.6M | Modern CNN | Pending |
| ConvNeXt-Small | 50.2M | Modern CNN | Pending |
| DINOv2-Small | 21M | ViT+LoRA | Pending |
| DINOv2-Base | 86M | ViT+LoRA | Pending |
Results
| Model | F1 Macro | Accuracy | Best Fold F1 |
|---|---|---|---|
| ResNet18 | — | — | — |
| ResNet50 | — | — | — |
(Results will be updated as training completes)
Structure
Each model directory contains:
serie1_ce/resnet18/
fold_0.json # Per-fold training history (loss, F1, accuracy per epoch)
fold_1.json
fold_2.json
fold_3.json
fold_4.json
best_fold_0.pt # Best model weights per fold
best_fold_1.pt
...
results.json # Aggregated metrics (mean±std across folds)
config.json # Training configuration
Serie 2 — Supervised Contrastive Loss (planned)
Same 6 architectures with SupCon loss (Khosla et al., 2020). Trains same-class proximity + different-class separation in embedding space.
Hardware
All models trained on NVIDIA Quadro T2000 (4GB VRAM) in a Podman container with GPU passthrough.
Download
from huggingface_hub import snapshot_download
# Download a specific model
snapshot_download(
repo_id="marcellorusso/orchid-ncd-models",
allow_patterns="serie1_ce/resnet18/*",
local_dir="models",
)
# Download everything
snapshot_download(
repo_id="marcellorusso/orchid-ncd-models",
local_dir="models",
)
Citation
If you use these models, please cite the accompanying thesis (details TBD).