ellariel/baesyan-opt-test

Implementation of Bayesian Optimization to find the optimal learning rate of a deep neural network (CV)

Implementation of Bayesian Optimization to find the optimal learning rate of a deep neural network (CV)

• Please look at _bo.ipynb on this repo for results.
• Direct link to Google Collab

Brief description

• Here I have implemented two solutions
  • "optimizer out of the box" (https://github.com/bayesian-optimization)
  • "optimizer from scratch"
• The code, results and figures are presented in the corresponding sections.
• Note that the presented solution is only a simple but working solution, it does not provide the best optimal learning rate, since it uses one learning epoch for the ResNet model.

Full task description

Objective: Find the optimal learning rate of a deep neural network for CV Optimizer:

• Implement Bayesian Optimization (BO) as a global optimization approach

  • use a Gaussian Process as a predictive model (you can use any existing GP library)
  • implement expected improvement (EI) as an acquisition function
  • use an optimization budget of 10 function evaluations

• For the deep neural network, please follow the specifications below:

  • A rather small ResNet (details can be chosen by you, but you should implement it on your own)
  • Optimizer: SGD
  • Deep learning framework: PyTorch or JAX
  • Dataset: Fashion-MNIST

• Plotting:

  • Starting with the second iteration of Bayesian Optimization, plot all observations, the posterior mean, the uncertainty estimate, and the acquisition function after each iteration.
  • See here for an exemplary plot (on a different task): link

• Programming language: Python; minimal coding standards:

  • PEP8
  • Doc-Strings
  • README with instructions how to run it