Tetraopes Beetle Flower Detection Model

Model Description

• Pre-trained by OpenAI on 400M image-text pairs (CLIP)
• Fine-tuned on ImageNet-12k and ImageNet-1k
• Input resolution: 336×336 pixels

Intended Use

Primary Use Case

• Identifying Tetraopes beetle observations on iNaturalist that show beetles on milkweed flowers/fruits
• Supporting ecological studies of beetle-plant associations
• Enabling large-scale analysis of host plant usage patterns
• Filtering iNaturalist photos for downstream analysis (e.g., host plant identification)

Intended Users

• Ecologists studying plant-insect interactions
• Entomologists researching Tetraopes biology
• Researchers conducting co-phylogenetic analyses
• Biodiversity scientists working with iNaturalist data

Out-of-Scope Use

• General flower detection in arbitrary contexts
• Detection of beetles other than Tetraopes
• High-stakes decision making
• Any application where misclassification could cause harm

Limitations

1. Taxonomic Scope: Trained specifically on Tetraopes beetles; may not generalize to other beetle genera
2. Plant Scope: Optimized for milkweed (Asclepias) flowers/fruits; performance on other plants not evaluated
3. Image Quality: Performance depends on image quality typical of iNaturalist photos
4. Confidence Calibration: Based on validation analysis, only predictions with >80% confidence are recommended for use
5. Small Training Set: Trained on only 121 manually labeled images; may not capture all variation
6. Class Imbalance: Training data is approximately balanced, but real-world distribution is unknown

Training Data

Dataset Summary

• Source: Manually labeled photos from iNaturalist observations
• Total Images: 119
  • With flowers (yes): 63
  • Without flowers (no): 56
• Train/Val Split: 80/20 stratified by class (seed=42)
  • Training: ~95 images
  • Validation: ~24 images
• Image Source: iNaturalist community observations

Training Images

• For observations with multiple photos, the first photo was used for training/validation.
• Original quality images (full resolution) were downloaded from iNaturalist and then resized to 336×336 pixels during training.

Training/Validation Images WITH Flowers (n=63)

Training/Validation Images WITHOUT Flowers (n=56)

Training Procedure

Hyperparameters

• Model: Vision Transformer Large (ViT-L/14)
  • Architecture: vit_large_patch14_clip_336.openai_ft_in12k_in1k
  • Patch size: 14×14
  • Input resolution: 336×336 pixels
• Training Strategy: Transfer learning with fine-tuning
  • Frozen epochs: 3 (train only classification head)
  • Fine-tuning epochs: 3 (train all layers)
  • Total epochs: 6
• Optimization:
  • Optimizer: AdamW (fastai default)
  • Learning rate: Determined via lr_find() - valley of loss curve
  • Learning rate schedule: One-cycle policy
  • Batch size: 8 (per GPU)
  • Weight decay: 0.01 (fastai default)
• Hardware:
  • 2× NVIDIA RTX A5000 GPUs (24GB each)
  • DataParallel training for faster convergence
• Data Augmentation:
  • Geometric:
    • Random rotation: ±10°
    • Random zoom: 1.0-1.3×
    • Random scale: 0.8-1.0×
    • Random warp: ±0.2
    • Horizontal flip: 50%
    • Vertical flip: 50%
  • Photometric:
    • Random lighting: ±0.4
  • Advanced:
    • MixUp augmentation (α=0.4) - blends image pairs during training
  • Normalization: ImageNet statistics (mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
• Class Balancing:
  • Weighted sampling during training
  • Sample weights calculated using normalized inverse frequency
  • Formula: weight = total_samples / (num_classes × class_count)

Training Procedure Details

1. Data Preparation:
   • Original quality images downloaded from iNaturalist (full resolution, typically ~2048px)
   • Images resized to 336×336 using squish method (no cropping)
   • Applied augmentation transforms
   • Normalized with ImageNet statistics
2. Transfer Learning:
   • Phase 1 (3 epochs): Froze all layers except classification head
   • Phase 2 (3 epochs): Unfroze all layers for end-to-end fine-tuning
   • Used MixUp augmentation throughout training
3. Validation:
   • Evaluated on held-out 20% validation set
   • No augmentation applied during validation
   • Metrics: accuracy, error rate

Evaluation

Validation Results

• Validation Accuracy: 100% (25/25 images correct)
• Validation Error Rate: 0%

          Predicted No  Predicted Yes
Actual No           12              0
Actual Yes           0             13

Large-Scale Prediction Results

• Predictions:
  • Flower detected (yes): 31,419 (41.7%)
  • No flower (no): 44,009 (58.3%)
  • Mean confidence: 84.0%
• High-Confidence Results (yes prediction with >80% confidence):
  • Photos: 21,307 (28.2%)
  • Unique observations: 15,916
  • Recommendation: Only use predictions with >80% confidence based on manual validation

Performance by Taxon

• "High-Conf Photos" = predictions with >80% confidence showing flowers
• "High-Conf Obs" = observations with at least one high-confidence flower photo
• Some taxa have limited observations; interpret results with caution

How to Use

Requirements

Loading the Model

Making Predictions

Batch Predictions

Model Card Authors

Citation

Additional Information

Framework Versions

• Python: 3.13.9
• PyTorch: 2.6.0
• FastAI: 2.8.5
• timm: 1.0.21

Model Size

• Parameters: ~304M (ViT-Large)
• Model file size: ~3.4 GB

Computational Requirements

• Training: 2× NVIDIA RTX A5000 (24GB) - ~30 minutes total
• Inference:
  • GPU recommended (any CUDA-capable GPU with 8GB+ VRAM)
  • CPU inference possible but slow (~1-2 seconds per image)
  • FP16 recommended for faster inference

License

Contact

Acknowledgments

• Training data from the iNaturalist community
• Pre-trained weights from OpenAI (CLIP) and timm library
• Training framework: FastAI