/SRResNet

SuperResolution via preactive ResNet

Primary LanguageLua

SRResNet

This is the impelmentation of SRResNet, a part of SRGAN "Ledig, Christian, et al. "Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network." arXiv preprint arXiv:1609.04802 (2016).".

Modification made from the paper:

  1. residual blocks --> preactivation residual blocks

  2. first and last conv. layer --> 9x9 conv. layer

Implementation was done in Torch7 supporting cuda/cudnn backend. (PASCAL Titan X)

###To Train Model 'th main.lua -nEpochs 350 -batchSize 16 -learningRate 1e-4' options in training the model can be changed by user's preference and can be found in 'opts.lua'.

###To Test Model 'th main.lua -testOnly true -resume checkpoints -testData Set5' options in testing the model can be found in 'opts.lua'.