A lasagne implementation of the MD-RNN proposed in the paper "Pixel Recurrent Neural Networks" by Aäron van den Oord, Nal Kalchbrenner and Koray Kavukcuoglu
This repository features:
- Bi-Diag-LSTM
- Masked convolutions (also color channel masking)
- !! Proper shift down by 1 row of the right Diag-LSTM
- Residual connections
- Networks weights for MNIST
- 256-way softmax and sigmoid output layers
- Dynamic generation of MNIST images in an updating
matplotlibwindow
TODO:
- ConvCNN
- Pyramid-RNN (Row-RNN)
- Train on CIFAR-10
The PixelRNN implementation can converge to a decent solution in a few epochs of training on MNIST.
Compatible with Python >= 3.3.
System dependencies:
sudo apt-get install graphviz
Python dependencies:
pip install -r requirements.txt
Images sampled (generated) from the Bi-Diag-LSTM:
Original images (left) and the same images with their lower half completed by the Bi-Diag-LSTM (right):
Three steps of the dynamic pixel-by-pixel generation of an image:
