This is the implementation of the Variational Ladder Autoencoder. Training on this architecture with standard VAE disentangles high and low level features without using any other prior information or inductive bias. This has been successful on MNIST, SVHN, and CelebA.
LSUN is a little difficult for VAE with pixel-wise reconstruction loss. However with another recently work we can generate sharp results on LSUN as well. This architecture serve as the baseline architecture for that model.
You need to have the following
- python 2 or 3 (More stable on python 2)
- tensorflow (tested on 0.12)
- numpy, scipy, matplotlib
You don't need to prepare the dataset for MNIST if you installed the examples that come with standard tensorflow distributions. If you can import tensorflow.examples.tutorials.mnist then simply run
python main.py --dataset=mnist
You should see results like this
Where if we sample one layer on a grid, and other layers randomly, we can see that the bottom layer represents stroke width, middle layer represents digit width, and top layer digit identity. Note that we cannot fully decouple them because digit identity is correlated with digit width (There is not really a '1' that is wide)
For SVHN download the dataset from http://ufldl.stanford.edu/housenumbers/
Download the three files train_32x32.mat, test_32x32.mat , extra_32x32.mat and place them into a folder and pass the absolute path of this folder as the db_path argument. Run
python main.py --dataset=svhn --db_path=/path/to/svhn/dataset
You should see decoupling like below
where we sample one layer randomly holding other layers fixed. The model has learned to disentangle color, digit variations, digit identity and general style.
For CelebA download the dataset from http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html
Download the aligned and cropped version, unzip into a folder, and pass the path of that folder as db_path argument. Note that your db_path folder should directly contain all the image files without any additional nesting. You folder should look like this
Now you can train by running
python main.py --dataset=celebA --db_path=/path/to/celebA/dataset
You should see decoupling like below
- To use another type of regularization use
--reg=[reg]. Currently support--reg=klfor KL(q(z|x)||p(z)) and--reg=mmdfor MMD(q(z), p(z)). - To use a particular GPU/GPUs add option
--gpus=[ids]such as--gpus=0,1to use GPU 0 and 1. In general for any of the recommended datasets the network generates good samples after running for about an hour on a single Titan X. The visualizations in the paper are produced after 1-2 days of training, but the improvement becomes marginal after the initial few hours. - To use other architectures other than default, use
--netname=[name]. For supported architectures please refer to code. The name is the unique identifier for a network configuration, and all related training log, visualizations, and checkpoint files, etc will be stored in the directory$pwd/model/netname. For example, to run visualization with tensorboard usetensorboard --logdir=[that directory]. - To also visualize the training process with a GUI window add
--use_gui. By default all plots will be stored to network directory, this will also plot them in a window in addition to that. - To change batch size, add
--batch_size=[size] - To visualize and plot the autoencoding reconstruction of the model, add
--plot_reconstruction - To add Gaussian and salt and pepper noise to perform denoise training add
--denoise_train - To control the number of batches before we visualize and make plots, use
--vis_frequency=[num_batch]