-
Session: https://360idev.com/sessions/machine-learning-on-ios-hands-on-with-ibm-watson-and-core-ml/
-
Workshop Based on: https://watson-developer-cloud.github.io/watson-vision-coreml-code-pattern/arduino/introduction.html
-
Shortened Session Link: http://ibm.biz/coremlworkshop
-
Image Downloads: http://ibm.biz/arduino_images
-
Demo: https://www.ibm.com/watson/services/visual-recognition/demo
- Tom Mariewicz
- Yacine Rezgui
IBM teammates who primarily work with Watson
Workshop Skill level: Beginner
We will build an image recognition app we can use offline. For working Arduino boards and ones with issues like burnt-out capacitors. This can help us troubleshoot bad components.
Written by the Authors
Last Spring, Apple partnered with IBM to bring Watson (IBM's visual recognition system) to iOS.
GPU required to use some services. But IBM now offers offline cloud one that's pay-by-usage.
Watson is a brand over many different types of products
- Collection of APIs for implementing AI/ML/DL
- Integration with wide range of products/services
- Security/privacy
- Build apps to use Watson models on iOS (even offline)
Just using the API will mean smaller app sizes. Developer chooses when to pull down models.
- Custom models
Identify shoe brands (Adidas vs Nike)
- Upload training data from real world, then download the updated model
Go to this website:
We'll use Carthage (>0.29)
Get up and take 10 images of arduino boards OR dowload them at ibm.biz/arduino_images
Make sure models images are done in different backgrounds. In one example, they tested ML on identifying Russian tanks versus American tanks. Because all the training photos of Russian tanks were taken during the day, it tended to classify any night photos as American tanks.
Negative training sets of similar items is useful. For instance, instead of Arduinos, get Raspberry Pis and similar looking items that are close enough for relevant training. Taking background-only images (like empty table, empty hand, etc) can also help the model know what to look for.