Skip to content

maxkvant/LinearizedNNs

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

LinearizedNNs

Here I apply Neural Kernels, a new approach to train neural networks.

The repository is inspired by the following papers:

  1. Wide Neural Networks of Any Depth Evolve as Linear Models Under Gradient Descent
  2. Neural Kernels without Tangents

Files

src/pytorch_impl/nns - neural networks in pytorch

src/pytorch_impl/estimators - estimators that can be trained:

  • SgdEstimator - estimator that trains a normal neural network via SGD
  • LinearizedSgdEstimator - estimator that trains a linearized neural network via SGD
  • MatrixExpEstimator - estimator that trains a linearized neural network via closed form solution, which turns out to be matrix exponentital

src/pytorch_impl/classifier_training - training procedure that trains an estimator and saves metrics on its way

notebooks - jupyter notebooks

papers - PDFs of papers I might refer in future

Legacy Files

notebooks/legacy/Wide_NNs_pytorch - my implementation of the training method from the paper.

notebooks/leagcy/Resnet18 - my attempt, yet unsuccessful, to achieve 94% accuracy on Cifar-10 via ResNet18.