2c6829/litecnn-pure-cpp

LiteCNNPro Model - Pure C++ Inference

Ultra-lightweight CNN model for dog breed classification

Model Details

• Model: LiteCNNPro (Pure C++ implementation)
• Parameters: 600K
• Classes: 120 (Stanford Dogs dataset)
• Input: 224×224 RGB images
• Framework: PyTorch (training) → Pure C++ (inference)
• Memory: 26MB total (4MB weights + 22MB runtime)

Architecture

Stem: Conv2D(3→32) + BatchNorm + ReLU6
Features: 7× Depthwise Separable Conv blocks
  - Block 0: 32→64 (stride 2)
  - Block 1: 64→128 (stride 2)
  - Block 2-3: 128→256 (stride 2)
  - Block 4-6: 256→512
  - SE (Squeeze-Excitation) attention in each block
Classifier: AdaptiveAvgPool → FC(512→256) → FC(256→120)

Usage

Download Model

wget https://huggingface.co/2c6829/litecnn-pure-cpp/resolve/main/model_weights.bin
wget https://huggingface.co/2c6829/litecnn-pure-cpp/resolve/main/breed_classes.json

Build and Run

# Clone the inference server
git clone https://github.com/stupidcoderJung/litecnn-pure-cpp
cd litecnn-pure-cpp

# Place model files
mv model_weights.bin weights/
mv breed_classes.json .

# Build
mkdir -p build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release
make -j4

# Run server
./litecnn_server --port 8080

API Example

# Health check
curl http://localhost:8080/health

# Predict
curl -X POST http://localhost:8080/predict \
  -F "image=@dog.jpg"

Response:

{
  "predictions": [
    {
      "class_id": 81,
      "score": 0.95,
      "breed_en": "Border collie",
      "breed_ko": "보더 콜리"
    }
  ]
}

Performance

Metric	Value
Memory (RSS)	26 MB
Binary Size	803 KB
Weights Size	4.0 MB
Inference Time	<100ms (CPU)

Comparison:

• PyTorch: 322 MB → 92% reduction ✅
• LibTorch: 130 MB → 80% reduction ✅
• ONNX Runtime: 102 MB → 75% reduction ✅

Files

• model_weights.bin (4.0 MB) - Model weights in binary format
• breed_classes.json (7.4 KB) - 120 dog breeds (English + Korean)
• extract_weights.py - PyTorch checkpoint → binary converter

Training

The model was trained on the Stanford Dogs dataset with:

• Optimizer: AdamW
• Learning rate: 1e-3
• Augmentation: Random flip, rotation, color jitter
• Epochs: 50
• Best validation accuracy: ~85%

License

MIT License

Citation

@software{litecnn_pure_cpp_2026,
  author = {LiteCNN Team},
  title = {LiteCNN Pure C++ Inference Server},
  year = {2026},
  url = {https://github.com/stupidcoderJung/litecnn-pure-cpp}
}

Contact

• Repository: https://github.com/stupidcoderJung/litecnn-pure-cpp
• Issues: https://github.com/stupidcoderJung/litecnn-pure-cpp/issues