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.
| Dataset | SRDenseNet Paper | SRDenseNet PyTorch |
|---|---|---|
| Set5 | 32.02 | 31.58 |
| Set14 | 28.50 | 28.36 |
| BSD100 | 27.53 | 27.38 |
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.