guanyuelee / midrae

Code for paper Improving Representation Learning in Autoencoders via Multidimensional Interpolation and Dual Regularizations.

MIDR-AE

Code for Improving Representation Learning in Autoencoders via Multidimensional Interpolation and Dual Regularizations, which is accepted by International Joint Conference of Artificial Intelligence (IJCAI) 2019.

MIDR-AE (right) introduces Multidimensional Interpolation and Dual Regularizations to improve latent representations in autoencoders.

Requirements

Before running MIDR-AE, you need python==3.5.6 and the following python packages:

• cudnn==7.1.2
• numpy==1.15.2
• scipy==1.1.0
• tensorflow==1.8.0

For your ease, my conda environment is exported as file. You can easily restore the environment by typing command:

conda env create -f environment.yaml

It would take you about 30 minutes to download all necessary python packages.

Preparing datasets

Specify the directory for saving your datasets by:

export AE_DATA=./Data

All the downloaded dataset would be save in './Data'.

Before running the model, you should prepare dataset by executing:

python create_datasets.py

It would take a long time. So I also upload all the necessary files to my Baidu Cloud, please download all the *.tfrecord and save in './Data'.

Running the code

There are multiple base models:

• Baseline
• Denoising
• Dropout
• Variational AE
• VQVAE
• Adversarial AE
• ACAI

Our proposed model are listed as follows:

• MI-AE
• LR-AE
• DR-ACAI
• MIDR-AE

Find *.sh in runs directory, we have all optimal hyperparameters for various models specified there. Just run:

bash runs/miae.sh
bash runs/lrae.sh
bash runs/dr_acai.sh
bash runs/midrae.sh

You should be able to achieve the reported accuracy of single-layer classifier and clustering resuls:

You can also check the model checkpoints and intermediate images in directory './TRAIN'.

Results

Some of the results of MIDR-AE are shown are follows:

MNIST-32:

SVHN-32:

Finally, if you have any questions about the code, please feel free to email: [email protected]