Bracelet Neural Network

Learning how to use Pytorch by making an image classifier for the knots in bracelet patterns.

Note

A "fun" experiment. Project has been archived.

Notes

• Custom dataset
  • Scraped 10 images from braceletbook.com
  • Convert to grayscale and cropped out each knot for a total of 1222 images
  • Use cv2 to display images one by one and press keys to label
  • 0 = ↙, 1 = ↘, 2 = ↲, 3 = ↳
  • Split 80% train, 10% validation, 10% test
• Custom model
  • 3 x (convolution > residual block > pooling), then global average and flatten, then 2 x (linear > relu), then softmax
  • residual block: 2 x (convolution > batch norm > relu)
  • Tried pre norm and post norm and did not seem to have much effect
• Pretraining
  • Torchvision MNIST set with 0.5 chance of being inverted
  • 5 epochs
• Training
  • Cross entropy loss
  • 10 epochs
  • Stochastic gradient descent optimizer with learning rate 0.01 and momentum 0.9
  • Exponential learning rate scheduler with gamma 0.9

Why this setup?

Your guess is as good as mine.

Does it work?

Not really.