WebML #1

TensorFlow.js

Server-side / Node.js benefit

1. Use TensorFlow SavedModel without conversion
2. Run larger models than client-side (GPU memory limits)
3. Code in 1 language
4. Performance -  C bindings, JIT boost for pre/post-processing

Backend

Client-side superpowers

1. privacy
2. low latency
3. lower cost
4. interactivity
5. Reach and Scale (GPU support on 84% of devices via WebGL)

3 ways to use

1. Pre-trained models
2. Transfer learning: Retrain existing models on your data
3. Custom models: User's own model

 

 

Selecting model: Inference speed (ms), File size (MB), Runtime RAM usage (MB)