insect-detection-vit

Insect Classification: Vision Transformer (ViT) Model

This model is a high-accuracy fine-tuned Vision Transformer (ViT) (google/vit-base-patch16-224-in21k backbone) trained to classify 21 distinct agricultural insect pest and beneficial species.

Training Hardware & Specs

• Compute: NVIDIA A100 Tensor Core GPU (SXM4 40GB)
• Total Training Steps: 16,600 iterations (20 full epochs)
• Training Time: 31 minutes 25 seconds
• Base Architecture: Vision Transformer Base (vit-base-patch16-224)
• Image Resolution: 224 x 224 px

---

Evaluation Benchmark & Training History

Top Metrics Achieved:

• Final Top-1 Accuracy: 98.51%
• Macro F1-Score: 98.58%
• Minimal Training Loss: 0.000100

---

Supported Taxa / Classes (21 Classes)

ant, aphid, bees, butterfly, caterpillar, cicada, dragonfly, grasshopper, green_lacewing, ladybug, leafhopper, mantis, mole_cricket, planthopper, rhino_beetle, rice_bug, spider, stem_borer, stink_bug, undefined, weevil

---

Quick Inference Usage

from transformers import AutoImageProcessor, AutoModelForImageClassification
from PIL import Image
import torch

model_id = "Mustafa5645344/insect-detection-vit"

# Load image processor & model
processor = AutoImageProcessor.from_pretrained("google/vit-base-patch16-224")
model = AutoModelForImageClassification.from_pretrained(model_id)

image = Image.open("insect_sample.jpg")
inputs = processor(images=image, return_tensors="pt")

with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.softmax(outputs.logits, dim=-1)
    predicted_idx = outputs.logits.argmax(-1).item()

predicted_label = model.config.id2label[str(predicted_idx)]
confidence = probs[0][predicted_idx].item()

print(f"Species: {predicted_label} (Confidence: {confidence*100:.2f}%)")