jojo-ai-mst/mm-deepsnake-v2

🐍 MM-DeepSnake v2 — Myanmar Snake Species Classifier

MM-DeepSnake is an AI-powered image classification model for identifying 10 snake species native to Myanmar. Built to support wildlife conservation, public safety, and biodiversity research, this model achieves 93.1% Top-1 accuracy and 99.7% Top-5 accuracy using a YOLOv2 medium classification backbone fine-tuned on a curated Myanmar snake image dataset.

🌐 Live Demo: mmdeepsnake.vercel.app

This is Version 2 — a major upgrade over the original VGG19-based v1 model (~62% val accuracy), delivering a +31% accuracy improvement.

မြန်မာနိုင်ငံတွင်တွေ့ရှိရသော မြွေမျိုးစိတ်များကို အမျိုးအစား ခွဲခြားရန်အတွက် ရေးဆွဲထားသော AI Project တစ်ခုဖြစ်သည်။

---

🔍 Model Details

Property	Details
Model Type	Image Classification
Architecture	YOLOv2m-cls (yolo26m-cls.pt)
Input Size	640 × 640 px
Output Classes	10 Myanmar snake species
Framework	Ultralytics 8.4.15 (PyTorch)
Parent Model	yolo26m-cls.pt (ImageNet pretrained)
Developed by	မင်းစည်သူ (Min Si Thu)
Website	mmdeepsnake.vercel.app
License	Apache 2.0

---

📊 Performance

Metric	Value
Top-1 Accuracy	93.14%
Top-5 Accuracy	99.74%
Training Loss (final)	0.00994
Epochs Trained	100
Training Time	7m 23s
Compute	RTX PRO 6000

Version Comparison

Version	Architecture	Input Size	Top-1 Accuracy	Top-5 Accuracy
v1	VGG19 (TensorFlow/Keras)	300 × 300	~62%	—
v2 (this model)	YOLOv2m-cls	640 × 640	93.14%	99.74%

---

🖼️ Sample Inference

Sample predictions from MM-DeepSnake v2 on unseen test images.

---

🐍 Supported Species (Classes)

The model classifies images into 10 snake species found across Myanmar, covering both venomous and non-venomous species common to the region. Visit the project website for the full species list with descriptions and venom information:

👉 mmdeepsnake.vercel.app/pages/snakes.html

---

🚀 How to Use

Requirements

pip install ultralytics>=8.4.19

Inference (Python)

from ultralytics import YOLO

# Load model
model = YOLO("jojo-ai-mst/mm-deepsnake-v2")  # or local path to model.pt

# Run inference
results = model("path/to/snake_image.jpg")

# Print top prediction
for result in results:
    top1_class = result.names[result.probs.top1]
    top1_conf = result.probs.top1conf.item()
    print(f"Predicted species: {top1_class} ({top1_conf*100:.1f}% confidence)")

    # Top-5 predictions
    for idx in result.probs.top5:
        print(f"  {result.names[idx]}: {result.probs.data[idx]*100:.1f}%")

Inference (CLI)

yolo classify predict model=jojo-ai-mst/mm-deepsnake-v2 source=snake_image.jpg

Batch Inference

from ultralytics import YOLO

model = YOLO("jojo-ai-mst/mm-deepsnake-v2")

results = model("path/to/images/", stream=True)

for result in results:
    print(result.path, "→", result.names[result.probs.top1])

---

🏋️ Training Details

Dataset

The model was trained on a private dataset of Myanmar snake images, collected from across the country and carefully labeled by species.

Split	Images
Train	1,300
Validation	201
Test	379
Total	~1,928

The training dataset is not publicly available to protect data integrity and prevent misuse.

Hyperparameters

Parameter	Value
Epochs	100
Batch Size	16
Image Size	640
Optimizer	Auto (SGD)
Initial LR (lr0)	0.01
Final LR (lrf)	0.01
Momentum	0.937
AMP	✅ Enabled
Dropout	0.0
Flip LR	0.5
Erasing	0.4
HSV-H / HSV-S / HSV-V	0.015 / 0.7 / 0.4
Close Mosaic	10
Pretrained	✅ Yes (ImageNet)

For further Training Command

yolo train device=2 model=yolo26m-cls.pt \
  data=your-dataset-path \
  project=min-si-thu/mm-deepsnake \
  name=experiment-1 \
  epochs=100 batch=16 imgsz=640

⚠️ Requires ultralytics>=8.4.19

Hardware & Environment

Property	Details
GPU	NVIDIA RTX PRO 6000
CPU	AMD EPYC 9655 (96-core)
Platform	Ultralytics Cloud (Docker)
OS	Linux 5.15.0 (Ubuntu, glibc 2.35)
Python	3.11.14
Ultralytics	8.4.15

---

⚠️ Limitations & Biases

• Dataset size: ~1,928 images across 10 classes is relatively small. Performance may degrade on unusual lighting, rare poses, or juvenile specimens.
• Geographic scope: Trained exclusively on Myanmar snake species. This model is not suitable for identifying snake species from other regions.
• No rejection mechanism: The model has no out-of-distribution detection. Passing non-snake images may yield confident but incorrect predictions.
• Class imbalance: Some species may have fewer training samples, potentially resulting in lower per-class accuracy for underrepresented species.
• Not a medical tool: This system is intended to assist — not replace — professional identification, especially in snakebite emergency contexts.

---

🌏 Intended Use

Recommended uses:

• Wildlife research and biodiversity surveys in Myanmar
• Public snake safety education and awareness
• Field identification assistance for researchers and conservationists
• Citizen science platforms and mobile applications

Out-of-scope uses:

• Medical diagnosis or snakebite treatment decisions
• Identification of non-Myanmar snake species
• Real-time safety-critical systems without human expert oversight

---

📖 Citation

If you use this model in your research or applications, please cite:

@misc{mm-deepsnake-v2-2026,
  author       = {မင်းစည်သူ (Min Si Thu)},
  title        = {MM-DeepSnake v2: Myanmar Snake Species Classification},
  year         = {2026},
  url          = {https://huggingface.co/jojo-ai-mst/mm-deepsnake-v2},
  note         = {YOLOv2m-cls fine-tuned on private MM Snakes dataset, Top-1 Acc: 93.1\%}
}

---

🔗 Related Links

Resource	Link
🌐 Website	mmdeepsnake.vercel.app
🐍 Snake Species Info	mmdeepsnake.vercel.app/pages/snakes.html
☠️ Venom Information	mmdeepsnake.vercel.app/pages/venom.html
📖 About the Project	mmdeepsnake.vercel.app/about.html
🤗 Author Profile	huggingface.co/jojo-ai-mst
🔧 Ultralytics Docs	docs.ultralytics.com

---

📄 License

This model is released under the Apache 2.0 License. See LICENSE for full terms.

The base model weights (yolo26m-cls.pt) are provided by Ultralytics under their respective license.

---

*Developed with ❤️ for Myanmar's wildlife and communities by မင်းစည်သူ