junbiaocui / grlf_gpg

A general representation learning framework with generalization performance guarantees.

Code of "A General Representation Learning Framework with Generalization Performance Guarantees"

1. Fold Structure

1.1 /GRLF_GPG/Core/

The core function modules, including:

(1) Radius.py

Code implementation of Algorithm 1 in Appendix B.1.

(2) Margin_DisConvexHull.py

Code implementation of Algorithm 2 in Appendix B.1.

(3) VC_Dim_KernelSelection.py

Code implementation of Algorithm 3 in Appendix B.2.

(4) VC_Dim_DNNBoostFramework.py

Code implementation of Algorithm 4 in Appendix B.3 and Fully-Connected Network.

(5) StandNet.py

Code implementation of Standardization Network, a part of Algorithm 4. See formula (12)-->(13).

(6) VC_Dim_Loss.py

Code implementation of VC dimension based loss, a part of Algorithm 4. See formula (14).

(7) DataSet.py

Code implementation of some data processing modules, including: load data from file and generate epoch, batch, etc.

1.2 Demos to illustrate how the proposed methods work,

including:

(1) Demo_KernelSelection.py

An experiment on Taichi data set, 628+629=1257 training samples, Gaussain kernel, 2000 kernel parameters.

(2) Demo_DNNBoost.py

An experiment on MNIST data set, 10x10=100 training samples, 10000 test sample, FCNet.
When VC_Dim_Loss_weight>0, the proposed boosting module will work.
When VC_Dim_Loss_weight=0, the proposed boosting module will not work.

1.3 /GRLF_GPG/DataSet/

Two data sets, including:

(1) /TaiChi/

Files of TaiChi data set.

(2) /MNIST/

Files of MNIST data set.

1.4 /GRLF_GPG/Result/

The default root folder of result files, including:

(1) /TaiChi/

Result files of experiments on Taichi data set.

(2) /MNIST/

Result files of experiments on MNIST data set.

1.5 ReadMe.md

This file.

2. Runtime Environment

2.1 OS (Operating System )

Ubuntu 18.04.5

2.2 IDE (Integrated Development Environment)

PyCharm 2020.1.5

2.3 Main Libraries

Python 3.8.12
PyTorch 1.7.0
Numpy 1.21.2
hdf5storage 0.1.16
scipy 1.7.1
etc.