Skip to content

Latest commit

 

History

History
49 lines (39 loc) · 2.23 KB

README.md

File metadata and controls

49 lines (39 loc) · 2.23 KB
Raw

In three clicks, make an iPhone app that runs image recognition on every video frame coming live from the camera

This app will create an iPhone app that uses Tensorflow image recognition, trained on your own images. You can make a trail mix calorie counter, a skin cancer detector, a cat in sofa detector, a cake recognizer, or whatever you can come up with.

Prepare

  • Install Xcode 8 or higher
  • Install Docker
  • Install Git
  • Make sure Docker and git are running
  • To make training faster, open Docker Preferences > Advanced, and increase CPUs and memory

Prepare your images and videos

  • Start Trainer
  • Click Browse
  • You can train your model with both images and videos
  • To train with images, go to images/originals
  • Create a new folder named after the object you want the app to recognize, e.g. "burgers"
  • Put pictures of burgers into the burgers folder
  • To train with videos, go to videos
  • Add your videos. Rename the video files e.g. burgers1.mov, burgers2.mov, etc

Start training

  • Click Start Training. The first training session will take about 30 minutes, depending on how many videos and images you have.

Make your app

  • Click the Xcode icon. Creating your app will take about 45 minutes.
  • Go to ~/projects/tf_files/tensorswift_ios/tensorswift
  • Open the xcode project and open Config.swift
  • Add your labels and which web page the app should open when it recognizes them

Add images to your model

  • Add the images to images/originals and videos
  • Click Start Training
  • In Xcode's left hand file panel, delete the files in the folder 'model'
  • Click Reveal model files and drag the files into the model folder in Xcode

Next steps

Feel free to grab one of these, or get in touch about any of them:

  • Optimize the model for inference
  • quantisize_graph for smaller model file size
  • Show Tensorboard vital stats directly in the app
  • Add UI for hyperparameters
  • Fix reset - it doesn't seem to do anything. Should remove bottlenecks and resize logs
  • Make it more clear how to retrain the model