RPMNet

Implement of our paper "Residual Refine based Pseudo Multi-frame Network for Efficient Single Image Super-Resolution"

Usage

Prepare

For more fair comparison with the state-of-art method using MATLAB, we use same script to do datasets generate and then export them with HDF5 file. So the model in PyTorch could receive same date as previous method done.

• For training, download DIV2K dataset and place the folder into prepare. Then run the generate_train.m using MATLAB. A big file in train.h5 will appear after it down.

• For testing, download Testing datasets, and change the variable folder to the place where test datasets in. Then run it in Matlab. A folder named testdatasets will appear.

• For real time loss monitor, we use tensorboardX

Training

usage: train.py [-h] [--batchSize BATCHSIZE] [--nEpochs NEPOCHS] [--lr LR]
                [--step STEP] [--cuda] [--start-epoch START_EPOCH]
                [--threads THREADS] [--momentum MOMENTUM]
                [--weight-decay WEIGHT_DECAY] [--pretrained PRETRAINED]
                [--dataset DATASET] [--number NUMBER]

RPMNet

optional arguments:
  -h, --help            show this help message and exit
  --batchSize BATCHSIZE
                        training batch size
  --nEpochs NEPOCHS     number of epochs to train for
  --lr LR               Learning Rate. Default=1e-4
  --step STEP           Sets the learning rate to the initial LR decayed by
                        momentum every n epochs
  --cuda                Use cuda?
  --start-epoch START_EPOCH
                        Manual epoch number (useful on restarts)
  --threads THREADS     Number of threads for data loader to use, Default: 1
  --momentum MOMENTUM   Momentum, Default: 0.9
  --weight-decay WEIGHT_DECAY, --wd WEIGHT_DECAY
                        weight decay, Default: 1e-4
  --pretrained PRETRAINED
                        path to pretrained model (default: none)
  --dataset DATASET     path to load dataset
  --number NUMBER       path to load dataset

Testing

usage: test.py [-h] [--cuda] [--model MODEL] [--imagepath IMAGEPATH]
               [--scale SCALE]

SR test

optional arguments:
  -h, --help            show this help message and exit
  --cuda                use cuda?
  --model MODEL         model path
  --imagepath IMAGEPATH
                        image path
  --scale SCALE