sca's Introduction

Scatter Component Analysis (SCA)

MATLAB implementation of Scatter Component Analysis for domain generalization proposed in paper

Ghifary, M., Balduzzi, D., Kleijn, W. B., & Zhang, M. (2017). Scatter component analysis: A unified framework for domain adaptation and domain generalization. IEEE transactions on pattern analysis and machine intelligence, 39(7), 1414-1430.

Getting Started

Prerequisites

The code is tested using MATLAB R2017b on Windows 10. Any later version should work normally.

Running the tests

In MATLAB, change your current folder to "SCA" and run one of the file demo.m to see whether it could run normally.

The file demo.m does the following:

Load synthetic data from "./syn_data/data.m";
Prepare source sample sets (put sample sets 1, 2 in a MATLAB cell array), validation set (sample sets 3, 4 in a matrix), and test set (sample set 5 in a matrix);
Learn transformations using SCA on the source sample sets and validate hyperparameters on the validation set.
Apply the optimal transformation on the test set.

Apply on your data

Usage

Change your current folder to "SCA" and use the following commands

[test_accuracy, predicted_labels, Zs, Zt] = SCA(X_s_cell, Y_s_cell, X_t, Y_t, params)

Description

Function SCA()

Input	Description
X_s_cell	cell of (n_s*d) matrix, each matrix corresponds to the instance features of a source domain
Y_s_cell	cell of (n_s*1) matrix, each matrix corresponds to the instance labels of a source domain
X_t	(n_t*d) matrix, rows correspond to instances and columns correspond to features
Y_t	(n_t*1) matrix, each row is the class label of corresponding instances in X_t
params	optional parameters, details can be found in SCA.m

Output	Description
test_accuracy	test accuracy on target instances
predicted_labels	predicted labels of target instances
Zs	projected source domain instances
Zt	projected target domain instances

Authors

Shoubo Hu - shoubo [dot] sub [at] gmail [dot] com

See also the list of contributors who participated in this project.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

Hat tip to Ya Li for his Conditional Invariant Domain Generalization (CIDG) code.

sca's People

Contributors

Stargazers

Watchers

sca's Issues

Question about the original paper and implementation

I don't understand your computing P and Q. Why the shapes of them are both (n_total, n_total)?

Questions about implementation

Hi,
Thx for a great implementation!

I have some questions about it compared to the paper:
File SCA_quantities.m:

Line 99:
I assume this implements calcultion of Ps (equation (13) in the paper).
Why there's no multiplication by nk?

Line 106 and onwards:
I assume this implements calculation of Qs (equation (14) in the paper).
The implementation looks very different from the equation.
Could you please explain?

Line 118 and onwards:
I assume D equals K*L*K in the paper (equation (20)).
The implementation looks very different from that.
Could you please explain?

Also, Algorithm 1 in the paper tells to first center the kernel at step 2 and then solve for B* in step 3. It follows, that D would need to be calculated with a centered kernel. But the code doesn't look like doing so (line 126).

Demo files, e.g. demo_2a_2a.m:
Line 69:
Why are the kernels centered with H?
In the paper, the only place where such centering is applied is in calculation of Qs in equation (14).
Please explain.

Thank you,
David

Recommend Projects

amber0309 / sca Goto Github PK