ameroyer / tf_deep_decoder

Reimplementation of the Deep Decoder architecture in Tensorflow

Deep Decoder

The code is based on the official Pytorch implementation of Deep Decoders available here.

The repository contains a Tensorflow implementation of Deep Decoders, as described in the paper Deep Decoder: Concise Image Representations from Untrained Non-convolutional Networks, Reinhard Heckel and Paul Hand, ICLR 2019.

Notes

• The code has been tested with Tensorflow 1.12 and Python 3.5+. Other dependecies include numpy, matplotlib and skimage (for reading/saving images only).
• The upsample_first argument in the decoder is inverted, to better match its meaning (i.e., when True, the upsampling operation occurs before the linear combination of channels)
• Currently the PSNR values are often slightly lower than the original implementation (possibly due to a different behavior in tf.image.resize_bilinear or some other error)
• LBFGS optimizer and weight decay are not implemented (but also not used in any of the present notebooks)