A implementation of ESRGAN using Tensorflow.
original paper: https://arxiv.org/abs/1809.00219
original implementation: https://github.com/xinntao/ESRGAN
see also my SRGAN implementation.
- Download vgg19.npy from here.
Put vgg19.npy in the folder where convert.py and vgg.py are located.
like this
...
│
├──convert.py
├──vgg.py
├──vgg19.npy
...
run convert.py.
python convert.py
after running, A vgg model dedicated to tensorflow will be output.
...
│
├── convert.py
├── vgg.py
├── vgg19.npy
├── modelvgg
│ ├── checkpoint
│ ├── model.ckpt-0
│ ├── model.ckpt-0
│ └── model.ckpt-0
...
- Download dataset from DIV2K dataset.
download the following four.
- Train Data Track 1 bicubic downscaling x4 (LR images)
- Train Data (HR images)
- Validation Data Track 1 bicubic downscaling x4 (LR images)
- Validation Data (HR images)
(If your PC has enough RAM, I recommend also The flickr2K data set proposed in EDSR.)
Put the downloaded dataset in the train_lr, train_hr, val_lr and val_hr folders. like this
...
│
├── train_lr
│ ├── 0001x4.png
│ ├── 0002x4.png
│ ├── ...
│ └── 0800x4.png
├── train_hr
│ ├── 0001.png
│ ├── 0002.png
│ ├── ...
│ └── 0800.png
├── val_lr
│ ├── 0801x4.png
│ ├── 0802x4.png
│ ├── ...
│ └── 0900x4.png
├── val_hr
│ ├── 0801.png
│ ├── 0802.png
│ ├── ...
│ └── 0900.png
├── main.py
├── model.py
...
- Train ESRGAN.
python main.py
- After training, inference can be performed.
Put the images you want to evaluate into a folder and run the following.
example:
...
│
├── main.py
├── pred.py
├── modelGAN
│ ├── checkpoint
│ ├── model.ckpt-100000
│ ├── model.ckpt-100000
│ └── model.ckpt-100000
├── test
│ ├── yourpic1.png
│ ├── favpic.jpg
│ ...
│ └── smallpic.png
...
python pred.py test
LR and HR images from jbhuang0604/SelfExSR.
leftmost: bicubic upsampling
Second from left: SRGAN(my implementation)
Third from left: ESRGAN(this implementation)
rightmost: original HR image
