What-Where autoencoder. Tensorflow

This project contains Tensorflow implementation of Stacked What-Where Auto-encoders. Implementation uses transposed convolutions provided by tensorflow and custom upsampling and unpooling code.

Note: As of now, unpooling code is not working with Tensorflow versions 1.13 and 2.0.0-alpha (the latest versions in pip) due to a bug in scatter_nd. It has been fixed in nightly versions and should be released with any next Tensorflow package. Please, use one of the next tensorflow versions:

$ pip install tf-nightly-2.0-preview 
$ pip install tf-nightly-gpu-2.0-preview
$ pip install tf-nightly # 1.14 preview
$ pip install tensorflow-gpu==1.12.0

Features

• [x] Unpooling layer (implemented by @Peepslee forum)
• [x] Fixed-position upsampling with zeros (Inverting Visual Representations with Convolutional Networks)
• [x] Fixed-position upsampling with element copies
• [ ] Custom network architecture interpreter as in original paper to configure network with a single line i.e. '(16)5c-(32)3c-Xp'

Outputs

Run tensorboard for visualization:

tensorboard --logdir=./tmp/

Picture above shows output images for original mnist image (left), decoding of what-where autoecoder (center), decoding of convolutional autoencoder with naive upsampling (right) while using stride=7. Picture repeat the experiment of the original paper.

Dependencies

• Python 3.5
• Tensorflow 1.0 with GPU support
• Numpy

pip3 install tensorflow-gpu numpy

Running model

Running learning script:

python WWAE.py --batch_size=128 --max_epochs=2 --pool_size=7
python WWAE_keras.py --batch_size=128 --max_epochs=2 --pool_size=7
python WWAE_tf2.0.py