DRRN-pytorch

This is an unofficial implementation of "Deep Recursive Residual Network for Super Resolution (DRRN)", CVPR 2017 in Pytorch. [Paper]

You can get the official Caffe implementation here.

Usage

Training

usage: main.py [-h] [--batchSize BATCHSIZE] [--nEpochs NEPOCHS] [--lr LR]
               [--step STEP] [--cuda] [--resume RESUME]
               [--start-epoch START_EPOCH] [--clip CLIP] [--threads THREADS]
               [--momentum MOMENTUM] [--weight-decay WEIGHT_DECAY]
               [--pretrained PRETRAINED]

optional arguments:
  -h, --help            Show this help message and exit
  --batchSize           Training batch size
  --nEpochs             Number of epochs to train for
  --lr                  Learning rate. Default=0.1
  --step                Learning rate decay, Default: n=5 epochs
  --cuda                Use cuda?
  --resume              Path to checkpoint
  --clip                Clipping Gradients. Default=0.01
  --threads             Number of threads for data loader to use Default=1
  --momentum            Momentum, Default: 0.9
  --weight-decay        Weight decay, Default: 1e-4
 --pretrained      Path to the pretrained model, used for weight initialization (default: none)

Evaluation

usage: eval.py [-h] [--cuda] [--model MODEL] [--dataset DATASET]
               [--scale SCALE]

PyTorch DRRN Evaluation

optional arguments:
  -h, --help         show this help message and exit
  --cuda             use cuda?
  --model MODEL      model path
  --dataset DATASET  dataset name, Default: Set5

An example of training usage is shown as follows:

python eval.py --cuda

Prepare Training dataset

• the training data is generated with Matlab Bicubic Interpolation, please refer Code for Data Generation for creating training files.

Performance

• We provide a rough pre-trained DRRN_B1U25 model trained on 291 images with data augmentation. The model can achieve a better performance with a smart optimization strategy. For the DRRN_B1U9 implementation, you can manually modify the number of recursive blocks here.

• The same adjustable gradient clipping's implementation as original paper.

• No bias is used in this implementation.

• No batch normalization is used in this implementation.

• Performance in PSNR on Set5

| Scale | DRRN_B1U25 Paper | DRRN_B1U25 PyTorch| | -------:| ----------------:| -----------------:| | x2 | 37.74 | 37.69 | | x3 | 34.03 | 34.02 | | x4 | 31.68 | 31.70 |