Skip to content

nanopony/keras-convautoencoder

Repository files navigation

Keras autoencoders (convolutional/fcc)

This is an implementation of weight-tieing layers that can be used to consturct convolutional autoencoder and simple fully connected autoencoder. It might feel be a bit hacky towards, however it does the job.

It requires Python3.x Why?.

Convolutional autoencoder [CAE] example

Run conv_autoencoder.py. Conv layer (32 kern of 3x3) -> MaxPool (2x2) -> Dense (10) -> DePool (2x2) -> DeConv layer (32 kern of 3x3)

Weights of Conv and Deconv layers are tied; MaxPool and DePool shares activated neurons.

ConvAutoEncoder MNIST representations

FCC autoencoder example

Run fcc_autoencoder.py. FСС (50) -> FСС (30) -> FСС (30) -> FСС (50) FСС MNIST representations

ConvAutoEncoder on 1100 cars

Vehicle images are courtesy of German Aerospace Center (DLR)

Remote Sensing Technology Institute, Photogrammetry and Image Analysis http://www.dlr.de/eoc/en/desktopdefault.aspx/tabid-5431/9230_read-42467/

Run 1100cars.py.

ConvAutoEncoder cars representations

About

Keras autoencoders (convolutional/fcc) [proof of concept]

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages