Skip to content

Drexel AI's Fall Term research project on efficiently searching for accurate neural net architectures

License

Notifications You must be signed in to change notification settings

drexelai/binary-search-in-neural-nets

Repository files navigation

binary-search-in-neural-nets

Drexel AI's Fall Term research project on efficiently searching for accurate neural net architectures

How to run

With python version over 3.6(I'm using 3.7.1), do

pip install -r requirements.txt

Then run

python run.py

There are some arguments that can be used. For example, to change epoch size, you can do

python run.py --epoch=200

The list of arguments can be seen at binary_search_networksbinary_search_parser.py

Current features

  • Gets data.
  • Preprocesses data.
  • Given input n, trains model. Prints out accuracy.
  • Tests model. Prints out accuracy.

Check list

Deadline: December 12, 2020

  • Pick a Dataset that can be generalized -> Using titanic dataset courtesy of https://www.openml.org/d/40945
  • Determine trendline over two end points, determine the slope and determine side to get rid of
  • Find the maximum value in a partially sorted array
  • Adapt the code so that user can pass n as input and run the entire pipeline (train + test + save)
  • Plot the loss and accuracy given a neural net for our problem
  • Add another dataset (churn model is added).
  • Implement linear search (O(N)) # need to train the model N times.
  • Implement binary search (O(log(N))) # need to train the model N times.
  • Determine trendline over two end points, determine the slope and determine side to get rid of.
  • Integrate the searches to pipeline.
  • Adapt the code so that user can pass n as input and run the entire pipeline (train + test + save).
  • Plot the loss and accuracy given a neural net for our problem.

About

Drexel AI's Fall Term research project on efficiently searching for accurate neural net architectures

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published