Theano implementation of the Diagonal BiLSTM and Pixel CNN models from Pixel Recurrent Neural Networks (van der Oord et al.) on binarized MNIST.
Not yet implemented:
- Residual connections
- 256-way softmax output
The Diagonal BiLSTM gets a dev-set NLL of 82.12 after 100K iterations (at which point it still hasn't fully converged).
Samples from the Diagonal BiLSTM after 100K iterations:
Ground-truth binarized MNIST samples:
