This code corresponding to the paper: Latent Space Factorisation and Manipulation via Matrix Subspace Projection (ICML2020).

The main website is here https://xiao.ac/proj/msp.

This code is based on

python 3.7
pytorch (version >= 1.4.0)
torchvision (version >= 0.4.1)

To train and test the model, you should download the CelebA dataset (from http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html).

You only need to put the two file: img_align_celeba.zip and list_attr_celeba.txt in the folder ./CelebA_Dataset/ .

Please run train_CelebA.py to train the model like:

> python3 train_CelebA.py

You can use the parameter -pg to show the training progress.

> python3 train_CelebA.py -pg

The trained model will be saved in ./model_save/ .

Alternatively, you can download the pre-trained model (from https://s3.eu-west-2.amazonaws.com/nn.models/MSP_CelebA.tch), and put the file MSP_CelebA.tch in ./model_save/ .

The file testing_CelebA.py can be used to generate the example pictures (including the picture used in the ICML paper).

> python3 testing_CelebA.py

The generated pictures will be in ./Outputs/ .

The code for the text experiment is being collated and will be released soon.

This work has been published in ICML2020. Here is the paper of the near camera-ready version. If you find MSP interesting, please consider citing:

  @incollection{icml2020_1832, author = {Li, Xiao and Lin, Chenghua and Li, Ruizhe and Wang, Chaozheng and Guerin, Frank}, booktitle = {Proceedings of Machine Learning and Systems 2020}, pages = {3211--3221}, title = {Latent Space Factorisation and Manipulation via Matrix Subspace Projection}, year = {2020} }  

This work is supported by the award made by the UK Engineering and Physical SciencesResearch Council (Grant number: EP/P011829/1).