Problem

Reducing Model Size

• Smaller bundle
• Smaller/faster downloads
• Reduced runtime memory usage

Core ML App Size

Resnet50

< 8 bit quantization

Obtaining Quantized Models

• Post-training quantization

• Train quantized
  - From scratch or re-training
  - Then convert quantized models to Core ML

Accuracy Tradeoff

Check Your Quantized Model Accuracy
with test data and metric relevant for your app

• Model dependent
• Use case dependent
• Active area of research

Demo

Sample Link

One Flexible Model

Combine Using Flexible Image Sizes

• One model
• No redundant code
• Faster model switching times

Flexibility Options

Which Models are Flexible?

Core ML Tools can check for you !

• Fully convolutional Neural Networks
  - Image processing
  - Object detection

Performance and Customization

For Loop vs Batch

Any Horses?

Custom Layers

Custom Models

Customization in Xcode

Customization