Invertible Image Rescaling | 环境搭建|简记

tech2023-02-25  126

作者:墨理 订阅号:墨理三生 博主声明:亲爱的童鞋,私信或者评论提出问题前,请关注我的订阅号墨理三生,打工人想恰饭哩,谢谢你的支持


如果您只是觉得我或者我的文章挺好看,也能来订阅号关注 墨理三生 ,小生在此鞠躬拜谢;


论文题目: Invertible Image Rescaling

论文链接 code链接

该github项目作者其实做了较为充分的工作,直接参考readMe即可顺利配置模型和数据,然后跑通代码; 本博文为博主简单环境搭建和测试简记,权供参考;


环境搭建:

服务器:ubuntu1~18.04 Quadro RTX 5000 16G CUDA版本 V10.0.130

conda create -n torch14 python=3.6.6 conda activate torch14 conda install pytorch==1.4.0 torchvision==0.5.0 cudatoolkit=10.0 -c pytorch pip install opencv-python pip install lmdb pyyaml pip install tb-nightly future —————————————————————————————— 根据代码需求,决定下面这几个是否安装: —————————————————————————————— pip install scipy pip install thop pip install scikit-image pip install tqdm

训练:

python train.py -opt options/train/train_IRN_x4.yml

测试:

python test.py -opt options/test/test_IRN_x2.yml

测试输出:

20-09-03 06:55:03.982 - INFO: Loading model for G [experiments/pretrained_models/IRN_x2.pth] ... Testing [set5]... 20-09-03 06:55:06.944 - INFO: ----Average PSNR/SSIM results for set5---- psnr: 39.786107 db; ssim: 0.969173. LR psnr: 39.535411 db; ssim: 0.980164. 20-09-03 06:55:06.944 - INFO: ----Y channel, average PSNR/SSIM---- PSNR_Y: 43.992743 dB; SSIM_Y: 0.987082. LR PSNR_Y: 46.446249 dB; SSIM_Y: 0.995682. Testing [set14]... 20-09-03 06:55:19.716 - INFO: ----Average PSNR/SSIM results for set14---- psnr: 36.694525 db; ssim: 0.948649. LR psnr: 37.053752 db; ssim: 0.970295. 20-09-03 06:55:19.716 - INFO: ----Y channel, average PSNR/SSIM---- PSNR_Y: 40.788999 dB; SSIM_Y: 0.977792. LR PSNR_Y: 44.050874 dB; SSIM_Y: 0.993581. Testing [BSD100]... 20-09-03 06:56:11.668 - INFO: ----Average PSNR/SSIM results for B100---- psnr: 38.793470 db; ssim: 0.982143. LR psnr: 37.263389 db; ssim: 0.968642. 20-09-03 06:56:11.668 - INFO: ----Y channel, average PSNR/SSIM---- PSNR_Y: 41.293070 dB; SSIM_Y: 0.987474. LR PSNR_Y: 44.887213 dB; SSIM_Y: 0.993612.

可能的排错记录:


FileNotFoundError: [Errno 2] No such file or directory: '../experiments/pretrained_models/IRN_x4.pth'


该代码 4倍 超分重建测试


对 Set5 数据集测试验证:

python test.py -opt options/test/test_IRN_x4.yml Testing [set5]... 21-01-27 02:33:10.778 - INFO: baby - PSNR: 35.333434 dB; SSIM: 0.933508; PSNR_Y: 37.732342 dB; SSIM_Y: 0.955487. LR PSNR: 39.276137 dB; SSIM: 0.978144; PSNR_Y: 46.773647 dB; SSIM_Y: 0.995603. 21-01-27 02:33:11.272 - INFO: bird - PSNR: 35.054365 dB; SSIM: 0.946841; PSNR_Y: 39.540321 dB; SSIM_Y: 0.976506. LR PSNR: 36.553521 dB; SSIM: 0.979724; PSNR_Y: 43.760511 dB; SSIM_Y: 0.996660. 21-01-27 02:33:11.700 - INFO: butterfly - PSNR: 30.891938 dB; SSIM: 0.942543; PSNR_Y: 33.096633 dB; SSIM_Y: 0.962123. LR PSNR: 35.061164 dB; SSIM: 0.991640; PSNR_Y: 42.038585 dB; SSIM_Y: 0.998511. 21-01-27 02:33:12.088 - INFO: head - PSNR: 31.252535 dB; SSIM: 0.788672; PSNR_Y: 35.056374 dB; SSIM_Y: 0.865289. LR PSNR: 38.004038 dB; SSIM: 0.969477; PSNR_Y: 45.692516 dB; SSIM_Y: 0.993992. 21-01-27 02:33:12.407 - INFO: woman - PSNR: 33.608163 dB; SSIM: 0.954969; PSNR_Y: 35.529980 dB; SSIM_Y: 0.966290. LR PSNR: 37.630199 dB; SSIM: 0.988383; PSNR_Y: 44.594586 dB; SSIM_Y: 0.997231. 21-01-27 02:33:12.441 - INFO: ----Average PSNR/SSIM results for set5---- psnr: 33.228087 db; ssim: 0.913306. LR psnr: 37.305012 db; ssim: 0.981474. 21-01-27 02:33:12.441 - INFO: ----Y channel, average PSNR/SSIM----

对 Set5 Set14 BSD100 Urban100 数据集测试验证:

python test.py -opt options/test/test_IRN_x4.yml Testing [set5]... 21-01-27 02:37:12.741 - INFO: ----Average PSNR/SSIM results for set5---- psnr: 33.229983 db; ssim: 0.913269. LR psnr: 37.305012 db; ssim: 0.981474. 21-01-27 02:37:12.742 - INFO: ----Y channel, average PSNR/SSIM---- PSNR_Y: 36.196054 dB; SSIM_Y: 0.945123. LR PSNR_Y: 44.571969 dB; SSIM_Y: 0.996399. Testing [set14]... 21-01-27 02:37:24.583 - INFO: ----Average PSNR/SSIM results for set14---- psnr: 30.118233 db; ssim: 0.861962. LR psnr: 35.531794 db; ssim: 0.969023. 21-01-27 02:37:24.584 - INFO: ----Y channel, average PSNR/SSIM---- PSNR_Y: 32.668332 dB; SSIM_Y: 0.901528. LR PSNR_Y: 42.321796 dB; SSIM_Y: 0.992785. Testing [B100]... 21-01-27 02:38:17.973 - INFO: ----Average PSNR/SSIM results for B100---- psnr: 29.937406 db; ssim: 0.863735. LR psnr: 36.085065 db; ssim: 0.964982. 21-01-27 02:38:17.973 - INFO: ----Y channel, average PSNR/SSIM---- PSNR_Y: 31.628820 dB; SSIM_Y: 0.882386. LR PSNR_Y: 43.187840 dB; SSIM_Y: 0.992275. Testing [Urban100]... 21-01-27 02:42:34.170 - INFO: ----Average PSNR/SSIM results for Urban100---- psnr: 29.239313 db; ssim: 0.893358. LR psnr: 34.777388 db; ssim: 0.965420. 21-01-27 02:42:34.170 - INFO: ----Y channel, average PSNR/SSIM---- PSNR_Y: 31.405625 dB; SSIM_Y: 0.915663. LR PSNR_Y: 41.283797 dB; SSIM_Y: 0.991604.

总结


可发现,就作者所用测试数据集,所公布模型的确可以达到所描述 PSNR、SSIM数值;

可见含金量之高,后续提升研究如果与这篇论文结果进行对比,还是很有难度的啊;

目测,这可能是2020年最强的一篇SISR研究的论文了;

脱坑了,脱坑了




最新回复(0)