Skip to content

We propose a lossless compression algorithm based on the NTK matrix for DNN. The compressed network yields asymptotically the same NTK as the original (dense and unquantized) network, with its weights and activations taking values only in {0, 1, -1} up to scaling.

Model-Compression/Lossless_Compression

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

61 Commits
 
 
 
 
 
 
 
 

Repository files navigation

NTK-LC

This repository contains code to reproduces the results in the paper "'Lossless' Compression of Deep Neural Networks: A High-dimensional Neural Tangent Kernel Approach" (NTK-LC) 1.

About the code

  • code/compression contains

    • code/compression/mnist (Experiment 2.1)
      • mnist_origin.py for classification with original dense neural network on MNIST dataset.
      • performance_match.py for classification with compressed neural network utlizing the proposed NTK*LC approach on MNIST dataset.
      • mnist_magnitude_pruning.py for classification with compressed neural network by magnitude pruning.
    • code/compression/cifar10 (Experiment 2.2)
      • vgg_net_cifar10.py for defining VGG19 suitable for CIFAR10
      • vgg_train.py for training VGG19 defined in vgg_net_cifar10.py on CIFAR10, and get parameters of convolution layers for feature extraction.
      • performance_origin.py for classification performance of the original dense neural network on CIFAR10.
      • performance_two.py for classification performance of the compressed neural network (compressed by our NTK-LC algorithm) on CIFAR10.
      • cifar10_magnitude_pruning.py for classification with compressed neural network by magnitude pruning.
  • code/spectral_characteristics (Experiment 1) contains

    • tilde_CK.py for verifying the consistency of spectrum distribution for theoretical (calculated with our theorem results) and practical (calculated by the original definition) conjugate kernel(CK).
    • plot_eigen.py for ploting eigenvalues and eigenvectors for a given matrix.
  • code/equation_solve contains

    • solve_equation.py for solving equatiosn to define parameters of activation functions
  • code/expect_cal contains

    • expect_calculate.py for expect calculated by numerical integration
    • expect_calculate_math.py for expect calculated with analytical expression
  • code/model_define contains

    • model.py for model defining
  • code/utils contains

    • activation_numpy.py for activations defined with numpy
    • activation_tensor.py for activations defined with torch
    • data_prepare.py for data preparation, containing data sampled from MNIST/CIFAR10 and generated GMM data
    • utils.py for some more utils

Dependencies

You can run following bash command to install packages used in this repository

pip install requirments.txt

or you can install follwing basic packages yourself:

Contact information

References

Footnotes

  1. Gu L, Du Y, Zhang Y, et al. Lossless Compression of Deep Neural Networks: A High-dimensional Neural Tangent Kernel Approach[J].link

About

We propose a lossless compression algorithm based on the NTK matrix for DNN. The compressed network yields asymptotically the same NTK as the original (dense and unquantized) network, with its weights and activations taking values only in {0, 1, -1} up to scaling.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages