[go: up one dir, main page]

Skip to content

Supplementary code for the paper "Stochastic Weight Matrix-based Regularization Methods for Deep Neural Networks" - an accepted paper of LOD2019

License

Notifications You must be signed in to change notification settings

rpatrik96/lod-wmm-2019

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Weight Matrix Modification - LOD2019 paper supplement

Authors: Patrik Reizinger & Bálint Gyires-Tóth

General description:

The project contains the source files (without the datasets) which implement WMM (Weight Matrix Modification,) a weight matrix-based regularization technique for Deep Neural Networks. In the following the proposed methods are shortly introduced, including the evaluation framework.

Weight shuffling

Weight shuffling is based on the assumption that locally the coefficients of a weight matrix are correlated. Based on this, we hypothesize that shuffling the weight within a rectangular window - which is under the beforementioned assumption a way of adding correlated noise to the weights - may help reduce overfitting.

Weight reinitialization

Weight reinitialization aims to reduce overfitting while partially reinitializing the weight matrix, thus in the case of a non-representative training set it may reduce the over-/underestimation of the significance regarding specific input data.

Usage:

The code can be run with typing the following command:

python ignite_main.py --model MODEL --dataset DS --num-trials TRIALS

Where MODEL can be one of following:

  • mnistnet
  • seqmnistnet
  • cifar10net
  • lstmnet
  • jsbchoralesnet

While DS should be (mismatch check is included in the code, the code structure was decided to be that way to enable the usage of multiple networks for the same dataset):

  • MNIST
  • CIFAR10
  • SIN (for the synthetic data) or SIN-NOISE (for noisy variant)

The default is to use Weight Shuffling, Weight Reinitialization can be selected by specifying --choose-reinit. TRIALS gives the number of runs by the hyper-optimization engine. The result of the hyper-optimization will be a .tsv file containing essential information about each training run.

More arguments concerning e.g. checkpointing or logging can be found in ignite_main.py.

The parameters of the optimizer (e.g. learning rate, momentum) can be set up in the ModelParameters class in descriptors.py.

Results

The 20 best results for each dataset and each method is included in the results directory, where the most important parameters are also included beside performance metrics.