EdgeCNN is a fully end-to-end machine learning project, covering the complete lifecycle from design and training to production-ready deployment
1. Model Architecture
- A custom edge-optimized convolutional neural network (CNN)
- Implements MobileNet-style depthwise separable convolution blocks for efficiency
2. Training
- Data Augmentation for better generalization
- Optimized with AdamW and cosine annealing learning rate scheduling
- Best model checkpointing
3. Evaluation & Benchmarking
- Quantitative comparison against established CIFAR-10 models
4. Production API
- A FastAPI service for model inference
- Image upload endpoint with class prediction with confidence
5. Deployment
- Containerized the application with Docker
- Designed for easy deployment to cloud providers (ex. pulling the image on an AWS EC2 instance)
Benchmarked locally on CPU, assuming $0.05 / hour computer cost:
| Model | Acc (%) | Params (M) | Size (MB) | Latency (ms) | $/1k Inferences |
|---|---|---|---|---|---|
| EdgeCNN (Ours) | 91.70 | 0.136 | 0.524 | 1.424 | 0.000020 |
| ResNet20 | 92.12 | 0.272 | 1.046 | 1.813 | 0.000025 |
| MobileNetV2_x0_5 | 92.41 | 0.700 | 2.743 | 6.113 | 0.000085 |
| ResNet56 | 94.25 | 0.856 | 3.281 | 4.458 | 0.000062 |
| MobileNetV2_x1_4 | 93.88 | 4.334 | 16.715 | 8.625 | 0.000120 |
Takeaway:
EdgeCNN achieves competitive accuracy while dramatically reducing inference time, memory footprint, and operational cost