Improved Depth Conrollable ARNet-Tensorflow

Tensorflow implemetation of depth conrllable ARNet (based on SNET)

Prerequisites

• python 3.6
• Tensorflow > 2.0
• Pillow
• cv2
• matplotlib
• argparse
• tqdm

Dataset

• Trainset : DIV2K
• Testset : LIVE1, BSD500
• dataset download link (무단입니다 private용으로만 공유)

Usage

For training

[from project root directory]
parser.add_argument("--gpu", type=str, default=0)  # -1 for CPU
parser.add_argument("--crop_size", type=list, default=[512, 512], nargs="+", help='Image size after crop.')
parser.add_argument("--buffer_size", type=int, default=20000, help='Data buffer size.')
parser.add_argument("--batch_size", type=int, default=16, help='Minibatch size(global)')
parser.add_argument("--patch_size", type=int, default=48, help='Minipatch size(global)')
parser.add_argument("--jpeg_quality", type=int, default=20, help='jpeg quallity')
parser.add_argument("--num_metrics", type=int, default=8, help='the number of metrics')
parser.add_argument("--num_filters", type=int, default=256, help='the number of filters')
parser.add_argument("--learning_rate", type=float, default=0.0001, help="lr")
parser.add_argument("--min_learning_rate", type=float, default=0.000001, help="min_lr")
parser.add_argument("--data_root_train", type=str, default="/projects/datasets/restoration/DIV2K/", help='Data root dir')
parser.add_argument("--data_root_test", type=str, default="/projects/datasets/restoration/LIVE1/", help='Data root dir')
parser.add_argument("--channels", type=int, default=3, help='Channel size')
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_file", type=str, default=None, help='file for resotration')
parser.add_argument("--graph_mode", type=bool, default=False, help='use graph mode for training')


python3 trainer.py [--gpu 0 --num_metrics 8 --exp_type 0] 

For Visualization (tensorboard)

[from project root directory]
tensorboard --logdir=./__outputs/summaries --port 8888

Result

References

• S-Net: A Scalable Convolutional Neural Network for JPEG Compression Artifact Reduction : https://arxiv.org/pdf/1810.07960.pdf

ToDo

Author

Dohyun Kim