Skip to content
/ bandit Public

Custom-made convolutional neural network lib (2022)

License

Notifications You must be signed in to change notification settings

ea-ae/bandit

Repository files navigation

Bandit

Experimenting with simple convolutional neural networks for image classification. All of the functionality was implemented from scratch and Eigen was used for the linear algebra.

Features

  • Stochastic gradient descent
    • Vectorization
    • Momentum
    • L2 regularization
    • Weight initialization
  • Layers: dense, convolutional
  • Cost functions: quadratic, cross-entropy
  • Activation functions: ReLu, leaky ReLu, sigmoid
  • Data loaders: MNIST, CIFAR-100, ImageNet

Performance

All the training-testing mini-batches were interleaved proportionally.

MNIST dataset:

  • 97.7% testing accuracy in 2 minutes over 3 epochs
  • 98.3% testing accuracy in 11 minutes over 15 epochs

Images

Screenshot

About

Custom-made convolutional neural network lib (2022)

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published