Skip to content

Latest commit

 

History

History
124 lines (91 loc) · 4.12 KB

README.md

File metadata and controls

124 lines (91 loc) · 4.12 KB
Raw

VaDE

This repository contains the Python implementation for our generative clustering method VaDE.

Further details about VaDE can be found in our paper:

Variational Deep Embedding : A Generative Approach to Clustering Requirements

  • Python-3.4.4
  • keras-1.1.0
  • scikit-learn-1.17.1

Replace keras/engine/training.py by training.py in this repository!!

(The modification version of keras/engine/training.py enables the simultaneous updating of the gmm parameters and the network parameters in our model.)

Usage

  • To train the VaDE model on the MNIST, Reuters, HHAR datasets:
python ./VaDE.py db

db can be one of mnist,reuters10k,har.

  • To achieve the 94.46% clustering accuracy on the MNIST dataset and generate the class-specified digits (Note that: unlike Conditional GAN, we do not use any supervised information during training):
python ./VaDE_test_mnist.py
  • To achieve the 79.38% clustering accuracy on the Reuters(685K) dataset:
cd $VaDE_ROOT/dataset/reuters
./get_data.sh
cd $VaDE_ROOT
python ./VaDE_test_reuters_all.py

Note: the data preprocessing code for the Reuters dataset is taken from (https://github.com/piiswrong/dec).

Face generation by VaDE on CelebA

(DCGAN-like network architecture)

  • Attribute-conditioned generation (sampled from each cluster) without using any supervised information

1-6 rows: 1.black/short hair, man; 2.black/long hair, woman; 3.gold/long hair, woman; 4.bald, sunglasses, man; 5.left side face, woman; 6.right side face, woman; image

  • Interpolation between cluster centers in latent space image

  • Vector arithmetic in latent space：right + left = front image

Run on Several Datasets (LYL)

Requirements

Python 3.5.6

requirements.txt (some of which need to be installed by pip)

Replace keras/engine/training.py with the training.py in this folder.

You can try this script:

conda create --name vade python=3.5.6 --yes
source activate vade
while read requirement; do conda install --yes $requirement || pip install $requirement; done < requirements.txt
cp training.py $CONDA_PREFIX/lib/python3.5/site-packages/keras/engine

HOWTO

Step 1

(You can skip this step for MNIST, Reuters10k, HAR, since the authors provided pretrained weights)

python VaDE.py dataset --mode pre-train

After training, you should see one json and one h5 added to pretrain_weights directory.

Step 2

python VaDE.py dataset

NOTE

  1. VaDE_test_*.py is not maintained (they just reproduce the last acc while running VaDE.py
  2. The authors may have fine-tuned the weights they provided. In contrast, running the code with steps 1-2 above yields much worse results on MNIST.
  3. There are two peculiar 1e-10 in VaDE.py. One appears in get_gamma function, which is not consistent with VaDE_test_*.py. If you want to maintain those testing scripts, make sure to ensure both or neither adds 1e-10 (otherwise testing acc will likely be significantly worse than training acc). The other appears in vae_loss function, but adding 1e-10 may deteriorate results on some datasets. The current strategy (adding 1e-10 only to MNIST, SVHN and CIFAR-100) is probably not optimal; you can modify line 76 to explore other possibilities.

Personal log(XSQ)

1e-10 seems to help avoid the model from cracking with p_c_z downflowing, yet the reason it works remains uncertified.

Duplicated experiment on all datasets worked well, yet results on fashion and stl were unstable.

Latest Best result(Accuracy):

mnist fashion-mnist stl-10 cifar-10 cifar-100 svhn
94.78 49.41 88.05 50.65 18.97 19.12