[Super Resolution] SRDenseNet - tensorflow implementation

tensorflow implementation of SRDenseNet

Prerequisites

python 3.x
Tensorflow > 1.x
Pillow
OpenCV
argparse

Properties (what's different from reference code)

This code requires Tensorflow. This code was fully implemented based on Python 3 differently from the original.
This code supports both RGB and YCBCR channel space.
This code supports tensorboard summarization
This code supports model saving and restoration
This code uses PILLOW library to resize image. Note that the performance of Bicubic function in PILLOW is lower than that of Matlab library.
Inference code should be changed to suit your environment.
This code is suitable for practical usage rather than research.
We use Resize-convolution instead Transpose-convolution. It is more flexible to handle various scale factor.

Usage

usage: python3 trainer.py --gpu 0 

[options]
""" system """
parser.add_argument("--exp_type", type=int, default=0, help='experiment type')
parser.add_argument("--gpu", type=str, default=1)  # -1 for CPU
parser.add_argument("--model_tag", type=str, default="default", help='Exp name to save logs/checkpoints.')
parser.add_argument("--checkpoint_dir", type=str, default='../__outputs/checkpoints/', help='Dir for checkpoints.')
parser.add_argument("--summary_dir", type=str, default='../__outputs/summaries/', help='Dir for tensorboard logs.')
parser.add_argument("--restore_model_file", type=str, default=None, help='file for restoration')
#parser.add_argument("--restore_model_file", type=str, default='../__outputs/checkpoints/SRDenseNet_SRDenseNet_model_default_10_11_00_55_14/model.ckpt-47000', help='file for resotration')
""" model """
parser.add_argument("--batch_size", type=int, default=16, help='Minibatch size(global)')
parser.add_argument("--patch_size", type=int, default=54, help='Minipatch size(global)')
#parser.add_argument("--patch_stride", type=int, default=13, help='patch stride') #we just sample patches randomly for simplicity
parser.add_argument("--operating_channel", type=str, default="RGB", help="operating channel [RGB, YCBCR")  # -1 for CPU
parser.add_argument("--num_channels", type=int, default=3, help='the number of channels')
parser.add_argument("--scale", type=int, default=3, help='scaling factor')
parser.add_argument("--data_root_train", type=str, default="./dataset/SR_training_datasets/T91", help='Data root dir')
parser.add_argument("--data_root_test", type=str, default="./dataset/SR_testing_datasets/Set5", help='Data root dir')