Machine learning

^{Unit 5: Machine Learning}

Day 1: Sketch Classifier

1.1 Experimenting with Sketch Classifier

• Visit this website to experiment with a trained machine learning algorithm.
• Trained Classifier Link: https://editor.p5js.org/kreier/sketches/a8qydXg7S (v1, v2)
• Play with the program and find what it does well and not so well, emphasis on the not-so-well. (Coding Challenge 158)
• Record a video that summarizes your findings and put it in the next page: 1.2

1.2 Sketch Classifier Observations

• After playing with the Sketch Classifier in 1.1, record a video where you show the strengths and the weaknesses of the classifier as it is now. What is it good at? What does it not do well?
• Submit your video at this Flipgrid link: https://flip.com/1e7ad802
• Last years submissions: https://flipgrid.com/4a7a2f4d
• To submit this assignment, take a screenshot of one of the biggest fails you observed while you played with this.

1.3 Training a Classifier

• (10:36) Students watch this video to learn how to use the code.
• Here is the folder of images for you to analyze in the task during class. Why does the classifier work as it does? How might you improve the data set to make it work better?
• Create Dataset for Classifier
  • Open the link using Google Chrome: This code was used to create the dataset (v1, v2)
  • Change the code based on your ideas for having better performance on triangles.
  • Run the code to make sure that the triangles are better than they were in the sample set.
  • Uncomment lines 31, 38 and 44. These will include the code to save the images of circles, squares, and triangles that appear. Then run the code for yourself.
  • Save all of the downloaded images into a single folder on your computer.
• Training the Classifier

  • Open the classifier training example: https://editor.p5js.org/kreier/sketches/nlmY0B2DH (v1) Download this example. You won’t be training on p5js website, instead on your local machine. Unpack the content of the downloaded zip-file.

  • Open the terminal, navigate to this downloaded folder and start your python3 webserver using the code python3 -m http.server 8000. If this gives an error, you first have to install the xcode command line tools with xcode-select -–install

  • Open a web browser in incognito mode and navigate to localhost:8000.

  • After training your classifier you should get 3 files: model_meta.json, model.json and model.weights.bin . But this one is based on only 30 images in /data.

  • Output left for the p5js website, output right on your local webserver:

  • Edit the sketch.js in the download folder, increase the input image list to 9 in line 18 to 9 and the epochs in line 46 to 60.

  • You have to stop the webserver with ctrl+c and start a new one on another port, for example on python3 -m http.server 8001 and then open localhost:8001 in the incognito browser

  • Copy your generated 300 images that you created into the training folder inside your downloaded example. Change the reference folder for the images to training in line 20, 21, 22

  • Train your model and download the 3 files model.json, model.weights.bin and model_meta.json

• Follow the instructions in this video to train your classifier on your computer:
• 7:38 AA - Shape Sketch Classifier - Training on your Computer

Day 2: Improving the Classifier

2.1 Testing the Classifier

• Open the classifier testing example: https://editor.p5js.org/kreier/sketches/jJXV1ycLp (v1)
• The included weights classify 19 to 20 correct: How well is your model doing?

Try this shape classifier - start this sketch and draw a triangle, circle or square into the field on the right.

Click classifier