Regularized_autoencoders(RAE)

This is the official implementation of the Paper titled 'From variational to deterministic Autoencoders'

If you find our work useful please cite us as the following.

@inproceedings{
ghosh2020from,
title={From Variational to Deterministic Autoencoders},
author={Partha Ghosh and Mehdi S. M. Sajjadi and Antonio Vergari and Michael Black and Bernhard Scholkopf},
booktitle={International Conference on Learning Representations},
year={2020},
url={https://openreview.net/forum?id=S1g7tpEYDS}
}

Watch a brief presentation

Full pdf

You can download a full PDF version of our paper from https://openreview.net/forum?id=S1g7tpEYDS

Set up

• Create a virtual environment virtualenv --no-site-packages <your_home_dir>/.virtualenvs/rae
• Activate your environment source <your_home_dir>/.virtualenvs/rae/bin/activate
• clone the repo git clone ...
• Navigate to RAE directory cd Regularized_autoencoders-RAE-
• Install requirements pip install -r requirements.txt
• Run training python train_test_var_reduced_vaes.py <config_id>

Data and pretrained models

###celebA

• Please download the celeba dataset
• Please pre process as desribed in the original paper. Centre crop 140X140 and resize to 64X64
• Place it under desired directory and specify the location in the dataloader.py file line 144
• the directory structure underneath must be as follows

celebA_root
    |____ train
            |___ train
                  |___ 0.png
                       1.png
                        .
                        .
                        .
    |____ test
            |___ test
                  |___ 182638.png
                       182639.png
                        .
                        .
                        .
    |____ val
            |___ val
                  |___ 162771.png
                       162772.png
                        .
                        .
                        .

CIFAR

• for CIFAR10 please prepare a .npz file cifar_10.npz
• this file should contain a numpy array of size tran_smaplesX32X32X3 under the key x_train
• and another numpy array of test samples of dimension 10000X32X32X3
• Values must be in the range 0-255
• please modify the path to root directory in line 143 in file dataloader.py

MNIST

• for MNIST please prepare a .npz file mnist_32x32.npz
• this file should contain a numpy array of size tran_smaplesX32X32X1 under the key x_train
• values must be in the range 0-255
• padd 2 columns and rows of zeros on all four sides to get the dimensionality of MNIST samples become 32X32X1
• modify the root directory path in line 125 in file dataloader.py

Checkpoints

• Please download the logs directory from this dropbox link
• specify the location to this directory in the config.py file
• and you are all set up to run

Misc

Example configurations

It is a big dictionar with primary and secondary entries. Primary entry holds all the common entries while the secondary entry is run specific entries. the run inde is simply the flat index number from top. So in the following example to run the setting under dictionary key 1 one must run with run_id 2 wile run_id 0 runs the first entry under dictionary key 0. All the results are generated in the log directory under sme directory hierarchy.

        configurations = \
            {0: [{'base_model_name': "rae"},
                                       {'expt_name': 'l2_regularization'},

                                       {'expt_name': 'l2_regularization'}
                   ],

             1: [{'base_model_name': "rae"},
                                       {'expt_name': 'spectral_normalization'}
                   ],

             2: [{'base_model_name': "rae"},
                                       {'expt_name': 'l2_regularization'},

                                       {'expt_name': 'l2_regularization'}
                   ],
            }