PixelCNN & PixelRNN in TensorFlow

TensorFlow implementation of Pixel Recurrent Neural Networks. This implementation contains:

Requirements

First, install prerequisites with:

$ pip install tqdm gym[all]

To train a pixel_rnn model with mnist data (slow iteration, fast convergence):

$ python main.py --data=mnist --model=pixel_rnn

To train a pixel_cnn model with mnist data (fast iteration, slow convergence):

$ python main.py --data=mnist --model=pixel_cnn --hidden_dims=64 --recurrent_length=2 --out_hidden_dims=64

To generate images with trained model:

$ python main.py --data=mnist --model=pixel_rnn --is_train=False

Samples generated with pixel_cnn after 50 epochs.

Below results uses two different parameters

[1] --hidden_dims=16 --recurrent_length=7 --out_hidden_dims=32
[2] --hidden_dims=64 --recurrent_length=2 --out_hidden_dims=64

Training results of pixel_rnn with [1] (yellow) and [2] (green) with epoch as x-axis:

Training results of pixel_cnn with [1] (orange) and [2] (purple) with epoch as x-axis:

Training results of pixel_rnn (yellow, green) and pixel_cnn (orange, purple) with hour as x-axis:

Taehoon Kim / @carpedm20