PyTorch SRDenseNet

Implementation of ICCV 2017 Paper: Image Super-Resolution Using Dense Skip Connections in PyTorch

Usage

Training

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

PyTorch DenseNet

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, Default: n=10
  --cuda                Use cuda?
  --resume RESUME       Path to checkpoint (default: none)
  --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)

An example of training usage is shown as follows:

python main.py --cuda

Evaluation

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

PyTorch SRDenseNet Eval

optional arguments:
  -h, --help         show this help message and exit
  --cuda             use cuda?
  --model MODEL      model path
  --dataset DATASET  dataset name, Default: Set5
  --scale SCALE      scale factor, Default: 4

An example of evalutation usage is shown as follows:

python eval.py --cuda

Demo

usage: demo.py [-h] [--cuda] [--model MODEL] [--image IMAGE] [--scale SCALE]

PyTorch SRDenseNet Test

optional arguments:
  -h, --help     show this help message and exit
  --cuda         use cuda?
  --model MODEL  model path
  --image IMAGE  image name
  --scale SCALE  scale factor, Default: 4

An example of demo usage is shown as follows:

python demo.py --model model_adam/model_epoch_60.pth --image butterfly_GT --scale 4 --cuda

Prepare Training dataset

• Please refer Code for Data Generation for creating training files.

• Data augmentations including flipping, rotation, downsizing are adopted.

Performance

• So far performance in PSNR is not as good as paper, since we only used 30,000 images for training while the authors used 50,000 images.

Misc.

• L1 Charbonnier loss from Deep Laplacian Pyramid Networks for Fast and Accurate Super-Resolution is applied instead of MSE loss, and PReLu is applied instead of ReLu due to gradient vanishing.