论文地址:https://arxiv.org/pdf/2108.10257.pdf
预训练模型下载:https://github.com/JingyunLiang/SwinIR/releases
训练代码下载:https://github.com/cszn/KAIR
测试代码:https://github.com/JingyunLiang/SwinIR
论文翻译:https://blog.csdn.net/hhhhhhhhhhwwwwwwwwww/article/details/124434886
测试:https://wanghao.blog.csdn.net/article/details/124517210
在写这边文章之前,我已经翻译了论文,讲解了如何使用SWinIR进行测试?
接下来,我们讲讲如何SwinIR完成训练,有于作者训练了很多任务,我只复现其中的一种任务。
地址:https://github.com/cszn/KAIR
这是个超分的库,里面包含多个超分的模型,比如SCUNet、VRT、SwinIR、BSRGGAN、USRNet等模型。
[外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-B5Md9i7H-1651410061139)(https://gitee.com/wanghao1090220084/cloud-image/raw/master/img/face_09_comparison.png)]
下载后解压,训练SwinIR的REANDME.md,路径:./docs/README_SwinIR.md
训练和测试集可以下载如下。 请将它们分别放在 trainsets
和 testsets
中。
任务 | 训练集 | 测试集 |
---|---|---|
classical/lightweight image SR | DIV2K (800 training images) or DIV2K +Flickr2K (2650 images) | set5 + Set14 + BSD100 + Urban100 + Manga109 download all |
real-world image SR | SwinIR-M (middle size): DIV2K (800 training images) +Flickr2K (2650 images) + OST (10324 images,sky,water,grass,mountain,building,plant,animal) SwinIR-L (large size): DIV2K + Flickr2K + OST + WED(4744 images) + FFHQ (first 2000 images, face) + Manga109 (manga) + SCUT-CTW1500 (first 100 training images, texts) |
RealSRSet+5images |
color/grayscale image denoising | DIV2K (800 training images) + Flickr2K (2650 images) + BSD500 (400 training&testing images) + WED(4744 images) | grayscale: Set12 + BSD68 + Urban100 color: CBSD68 + Kodak24 + McMaster + Urban100 download all |
JPEG compression artifact reduction | DIV2K (800 training images) + Flickr2K (2650 images) + BSD500 (400 training&testing images) + WED(4744 images) | grayscale: Classic5 +LIVE1 download all |
我下载了DIV2K数据集和 Flickr2K数据集,DIV2K大小有7G+,Flickr2K约20G。如果网速不好建议只下载DIV2K。
注:在选用classical任务,做训练时,只能使用DIV2K或者Flickr2K,不能把两种数据集放在一起训练,否则就出现维度对不上的情况,如下图:
暂时没有找到原因。
构建测试集,测试集的路径如下图:
由于表格中的测试集放在google,我不能下载,但是SwinIR的测试代码中有测试集,代码链接:https://github.com/JingyunLiang/SwinIR,下载下来直接复制到testsets文件夹下面。
构建训练集,将下载下来的DIV2K解压。将DIV2K_train_HR复制到trainsets文件夹下面,将其改为trainH。
将DIV2K_train_LR_bicubic文件夹的X2文件夹复制到trainsets文件夹下面,然后将其改名为trainL。
到这里,数据集部分就完成了,接下来开始训练。
首先,打开options/swinir/train_swinir_sr_classical.json文件,查看里面的内容。
"task": "swinir_sr_classical_patch48_x2"
训练任务的名字。
"gpu_ids": [0,1]
选择GPU的ID,如果只有一快GPU,改为 [0]。如果有更多的GPU,直接往后面添加即可。
"scale": 2 //2,3,48
放大的倍数,可以设置为2、3、4、8.
"datasets": {
"train": {
"name": "train_dataset" // just name
, "dataset_type": "sr" // "dncnn" | "dnpatch" | "fdncnn" | "ffdnet" | "sr" | "srmd" | "dpsr" | "plain" | "plainpatch" | "jpeg"
, "dataroot_H": "trainsets/trainH"// path of H training dataset. DIV2K (800 training images)
, "dataroot_L": "trainsets/trainL" // path of L training dataset
, "H_size": 96 // 96/144|192/384 | 128/192/256/512. LR patch size is set to 48 or 64 when compared with RCAN or RRDB.
, "dataloader_shuffle": true
, "dataloader_num_workers": 4
, "dataloader_batch_size": 1 // batch size 1 | 16 | 32 | 48 | 64 | 128. Total batch size =4x8=32 in SwinIR
}
, "test": {
"name": "test_dataset" // just name
, "dataset_type": "sr" // "dncnn" | "dnpatch" | "fdncnn" | "ffdnet" | "sr" | "srmd" | "dpsr" | "plain" | "plainpatch" | "jpeg"
, "dataroot_H": "testsets/Set5/HR" // path of H testing dataset
, "dataroot_L": "testsets/Set5/LR_bicubic/X2" // path of L testing dataset
}
}
上面的参数是对数据集的设置。
“H_size”: 96 ,HR图像的大小,和下面的img_size有对应关系,大小设置为img_size×scale。
“dataloader_num_workers”: 4,CPU的核数设置。
“dataloader_batch_size”: 32 ,设置训练的batch_size。
dataset_type:sr,指的是数据集类型SwinIR。
"netG": {
"net_type": "swinir"
, "upscale": 2 // 2 | 3 | 4 | 8
, "in_chans": 3
, "img_size": 48 // For fair comparison, LR patch size is set to 48 or 64 when compared with RCAN or RRDB.
, "window_size": 8
, "img_range": 1.0
, "depths": [6, 6, 6, 6, 6, 6]
, "embed_dim": 180
, "num_heads": [6, 6, 6, 6, 6, 6]
, "mlp_ratio": 2
, "upsampler": "pixelshuffle" // "pixelshuffle" | "pixelshuffledirect" | "nearest+conv" | null
, "resi_connection": "1conv" // "1conv" | "3conv"
, "init_type": "default"
}
upscale:2,放大的倍数,和上面的scale参数对应。
img_size:48,这里可以设置两个数值,48和64。和测试的training_patch_size参数对应。
官方提供的指令是基于DDP方式,比较复杂一下,好处是速度快。如下:
# 001 Classical Image SR (middle size)
python -m torch.distributed.launch --nproc_per_node=8 --master_port=1234 main_train_psnr.py --opt options/swinir/train_swinir_sr_classical.json --dist True
# 002 Lightweight Image SR (small size)
python -m torch.distributed.launch --nproc_per_node=8 --master_port=1234 main_train_psnr.py --opt options/swinir/train_swinir_sr_lightweight.json --dist True
# 003 Real-World Image SR (middle size)
python -m torch.distributed.launch --nproc_per_node=8 --master_port=1234 main_train_psnr.py --opt options/swinir/train_swinir_sr_realworld_psnr.json --dist True
# before training gan, put the PSNR-oriented model into superresolution/swinir_sr_realworld_x4_gan/models/
python -m torch.distributed.launch --nproc_per_node=8 --master_port=1234 main_train_psnr.py --opt options/swinir/train_swinir_sr_realworld_gan.json --dist True
# 004 Grayscale Image Deoising (middle size)
python -m torch.distributed.launch --nproc_per_node=8 --master_port=1234 main_train_psnr.py --opt options/swinir/train_swinir_denoising_gray.json --dist True
# 005 Color Image Deoising (middle size)
python -m torch.distributed.launch --nproc_per_node=8 --master_port=1234 main_train_psnr.py --opt options/swinir/train_swinir_denoising_color.json --dist True
# 006 JPEG Compression Artifact Reduction (middle size)
python -m torch.distributed.launch --nproc_per_node=8 --master_port=1234 main_train_psnr.py --opt options/swinir/train_swinir_car_jpeg.json --dist True
我没有使用上面的方式,而是选择用DP的方式,虽然慢一点,但是简单,更稳定。
在Terminal里面输入:
python main_train_psnr.py --opt options/swinir/train_swinir_sr_classical.json
即可开始训练。
运行结果如下:
等待训练完成后,我们使用测试代码测试。将模型复制到./model_zoo/swinir文件夹下面
输入命令:
python main_test_swinir.py --task classical_sr --scale 2 --training_patch_size 48 --model_path model_zoo/swinir/45000_G.pth --folder_lq testsets/Set5/LR_bicubic/X2
然后在result下面可以看到测试结果。
https://download.csdn.net/download/hhhhhhhhhhwwwwwwwwww/85258387