LeverX is happy to participate in a brand-new Swift Conference in Wroclaw to share our experience. The speaker from LeverX is Vitali Usau, Head of Innovation and Mobile Development.

LeverX at the plSwift Conference in Wroclaw

LeverX is happy to participate in a brand-new Swift Conference in Wroclaw to share our experience. The speaker from LeverX is Vitali Usau, Head of Innovation and Mobile Development.

PlSwift is an excellent opportunity to meet experts and talk on the most relevant topics. After two years of mostly online events, the organizators will be bringing back the joy of on-site meetups in a yet-to-be-discovered region of Poland.

 Vitali is going to speak on “Possible ways of building and running ML models on iOS devices”.

As the bonus for all participants, Vitali has provided useful links and key notes. Add to your bookmarks!

Key Notes and Useful Links

Where should I run the model?

Edge Сomputing. Advantages:

- Data Privacy (personal data (messages), IDs, financial data)
- Low latency (live video update, ~10ms)
- Cost ($0 :) )
- Ability to work offline

Cloud Computing. Advantages:

- Single place for model updating
- Platform/device agnostic
- No versioning
- Less hardware demanding

Cross-platform

- Tensorflow lite
- PyTorch Mobile

iOS specific

- Metal Performance Shaders
- CoreML (Create ML, Turi Create, coremltools)
- Other (Accelerate framework, Bender)

Other materials

• Core ML Survival Guide book
• Collection of pretrained Core ML models
• What do we know about ANE

Image classification notes:

• Create ML training is fast. It uses predefined feature extractor and trains only classifier
  
• Create ML can compete with custom models on small amount of data
• Create ML doesn’t require any data science skills
• Custom models allow you to get much higher score on a big amount of data
• Custom models allow control of the model size and accuracy
• Custom models are not limited to data augmentation techniques

Audio classification notes:

• Create ML is supposed to be trained on one class per data sample
• There are no augmentations
• Constant sampling rate (16 kHz), which limits the highest signal frequency at 8kHz
• Windows size is fixed (0.975 s), which gives poor timing resolution
• The sum of all probabilities is 1.0

Interested in software development? Follow the link to find the hot vacancies and join the team of a successful international IT company.