Multi-task Autoencoders for Domain Generalization

==========

This code contains the Python implementation of the Multitask Autoencoder (MTAE) algorithm based on the following paper:

M. Ghifary, W. B. Kleijn, M. Zhang, D. Balduzzi.
Domain Generalization for Object Recognition with Multi-task Autoencoders,
accepted in International Conference on Computer Vision (ICCV 2015), Santiago, Chile.
[pre-print]

Please cite the above paper when using this code.

Notes:

• This version is based on the Theano wrapper called keras.io and was written after the paper finished. The prior code of this work was implemented in MATLAB.
• Currently only works for the provided MNIST dataset (with 6 rotated views)
• Still not well commented

For questions and bug reports, please send me an email at mghifary[at]gmail.com.

Prerequisites

1. The following frameworks/libraries must be installed:
2. Python (version 2.7 or higher)
3. Numpy (e.g. pip install numpy)
4. Theano
5. Keras
6. Clone this repository, e.g.: git clone https://github.com/ghif/mtae.git
7. Run the main program to reproduce either Figure 4(c) or (d): ./run_mtae_gpu.sh
   • if you have a GPU, make sure that the nvcc compiler path is included in the environment variables.

TO DO :

• RAND-SEL procedure
• SVM classification
• Supervised Finetuning