Stacked what where autoencoders in tensorflow and keras.
mnist_swwae-tf.py, tensorflow compatible version of mnist_swwae.py from https://github.com/fchollet/keras/blob/master/examples/mnist_swwae.py
mnist_swwae-pure-tf.py, same as above but uses pure tensorflow, no keras.
mnist_swwae-pure-tf-paper.py, uses pure tensorflow but implements L2M error terms not computed in the above better matching derivation in paper [1].
ChannelNormPooling - this does max-pooling but pools based on the channel-wise norm. Thus, entire pixels with all their channels are pooled instead of pooling each channel independently.
FFT2D and IFFT2D layers. batchwise 2D FFT layers for image processing.
Schlick dynamic range compression method with auto brightness algorithm. [2]
[1] Stacked What-Where Auto-encoders Junbo Zhao, Michael Mathieu, Ross Goroshin, Yann LeCun https://arxiv.org/abs/1506.02351
[2] @incollection{schlick1995quantization, title={Quantization techniques for visualization of high dynamic range pictures}, author={Schlick, Christophe}, booktitle={Photorealistic Rendering Techniques}, pages={7--20}, year={1995}, publisher={Springer} }