xpixelgroup / basicsr

Open Source Image and Video Restoration Toolbox for Super-resolution, Denoise, Deblurring, etc. Currently, it includes EDSR, RCAN, SRResNet, SRGAN, ESRGAN, EDVR, BasicVSR, SwinIR, ECBSR, etc. Also support StyleGAN2, DFDNet.

English | 简体中文

⚡HowTo | 🔧Installation | 💻Training Commands | 🐢DatasetPrepare | 🏰Model Zoo

📕中文解读文档 | 📊 Plot scripts | 📝Introduction | | ⏳TODO List | ❓FAQ

🚀 We add BasicSR-Examples, which provides guidance and templates of using BasicSR as a python package. 🚀
📢 技术交流QQ群：320960100   入群答案：互帮互助共同进步
🧭 入群二维码 (QQ、微信)    入群指南 (腾讯文档)

---

BasicSR (Basic Super Restoration) is an open-source image and video restoration toolbox based on PyTorch, such as super-resolution, denoise, deblurring, JPEG artifacts removal, etc.
BasicSR (Basic Super Restoration) 是一个基于 PyTorch 的开源 图像视频复原工具箱, 比如 超分辨率, 去噪, 去模糊, 去 JPEG 压缩噪声等.

🚩 New Features/Updates

• ✅ July 26, 2022. Add plot scripts 📊Plot.
• ✅ May 9, 2022. BasicSR joins XPixel.
• ✅ Oct 5, 2021. Add ECBSR training and testing codes: ECBSR.

  ACMMM21: Edge-oriented Convolution Block for Real-time Super Resolution on Mobile Devices

• ✅ Sep 2, 2021. Add SwinIR training and testing codes: SwinIR by Jingyun Liang. More details are in HOWTOs.md
• ✅ Aug 5, 2021. Add NIQE, which produces the same results as MATLAB (both are 5.7296 for tests/data/baboon.png).
• ✅ July 31, 2021. Add bi-directional video super-resolution codes: BasicVSR and IconVSR.

  CVPR21: BasicVSR: The Search for Essential Components in Video Super-Resolution and Beyond

• More

---

If BasicSR helps your research or work, please help to ⭐ this repo or recommend it to your friends. Thanks😊
Other recommended projects:
▶️ Real-ESRGAN: A practical algorithm for general image restoration
▶️ GFPGAN: A practical algorithm for real-world face restoration
▶️ facexlib: A collection that provides useful face-relation functions.
▶️ HandyView: A PyQt5-based image viewer that is handy for view and comparison.
▶️ HandyFigure: Open source of paper figures
_{(ESRGAN, EDVR, DNI, SFTGAN) (HandyCrawler, HandyWriting)}

---

⚡ HOWTOs

We provide simple pipelines to train/test/inference models for a quick start. These pipelines/commands cannot cover all the cases and more details are in the following sections.

GAN
StyleGAN2	Train	Inference
Face Restoration
DFDNet	-	Inference
Super Resolution
ESRGAN	TODO	TODO	SRGAN	TODO	TODO
EDSR	TODO	TODO	SRResNet	TODO	TODO
RCAN	TODO	TODO	SwinIR	Train	Inference
EDVR	TODO	TODO	DUF	-	TODO
BasicVSR	TODO	TODO	TOF	-	TODO
Deblurring
DeblurGANv2	-	TODO
Denoise
RIDNet	-	TODO	CBDNet	-	TODO

✨ Projects that use BasicSR

• Real-ESRGAN: A practical algorithm for general image restoration
• GFPGAN: A practical algorithm for real-world face restoration

If you use BasicSR in your open-source projects, welcome to contact me (by email or opening an issue/pull request). I will add your projects to the above list 😊

📜 License and Acknowledgement

This project is released under the Apache 2.0 license.
More details about license and acknowledgement are in LICENSE.

🌏 Citations

If BasicSR helps your research or work, please cite BasicSR.
The following is a BibTeX reference. The BibTeX entry requires the url LaTeX package.

@misc{basicsr,
  author =       {Xintao Wang and Liangbin Xie and Ke Yu and Kelvin C.K. Chan and Chen Change Loy and Chao Dong},
  title =        {{BasicSR}: Open Source Image and Video Restoration Toolbox},
  howpublished = {\url{https://github.com/XPixelGroup/BasicSR}},
  year =         {2022}
}

Xintao Wang, Liangbin Xie, Ke Yu, Kelvin C.K. Chan, Chen Change Loy and Chao Dong. BasicSR: Open Source Image and Video Restoration Toolbox. https://github.com/xinntao/BasicSR, 2022.

📧 Contact

If you have any questions, please email [email protected], [email protected].

• QQ群: 扫描左边二维码 或者 搜索QQ群号: 320960100   入群答案：互帮互助共同进步
• 微信群: 我们的一群已经满500人啦，二群也超过200人了；进群可以添加 Liangbin 的个人微信 (右边二维码)，他会在空闲的时候拉大家入群~

 

(start from 2022-11-06)

basicsr's Issues

SRGAN testing configuration

How to write configuration for SRGAN(Not ESRGAN) testing? I always encounter the "missing key in state_dict" problem.

Dataloader / Prepearing Dataset

When I try to use urban100 with SFTGAN, I get the following error:

Traceback (most recent call last):
File "train.py", line 157, in
main()
File "train.py", line 71, in main
for i, train_data in enumerate(train_loader):
File "/usr/local/lib/python3.5/dist-packages/torch/utils/data/dataloader.py", line 286, in next
return self._process_next_batch(batch)
File "/usr/local/lib/python3.5/dist-packages/torch/utils/data/dataloader.py", line 307, in _process_next_batch
raise batch.exc_type(batch.exc_msg)
FileNotFoundError: Traceback (most recent call last):
File "/usr/local/lib/python3.5/dist-packages/torch/utils/data/dataloader.py", line 57, in _worker_loop
samples = collate_fn([dataset[i] for i in batch_indices])
File "/usr/local/lib/python3.5/dist-packages/torch/utils/data/dataloader.py", line 57, in
samples = collate_fn([dataset[i] for i in batch_indices])
File "/home/onurctirtir/Desktop/BasicSR/codes/data/LRHR_seg_bg_dataset.py", line 58, in getitem
seg = torch.load(HR_path.replace('/img/', '/bicseg/').replace('.png', '.pth'))
File "/usr/local/lib/python3.5/dist-packages/torch/serialization.py", line 301, in load
f = open(f, 'rb')
FileNotFoundError: [Errno 2] No such file or directory: '/home/onurctirtir/Desktop/datasets/urban100/img_002.pth'

Probably something is going wrong with me. A detailed description for "preparing dataset" would be super.

How did you get the ESRGAN pre training model?

How did you get the ESRGAN pre training model?

Some concerns on usage

I don't know if opening an issue for this is appropriate or not but I stuck around at some point:

1. In SFTGAN.json, what is expected for "datasets:dataroot_HR_bg:" ? What 'bg' corresponds for ?
2. If I'm not wrong, I should put HR images under 'img/' at some directory and '.pth' files under 'bicseg/' being in the same directory with 'img/'

Thanks in advance

Windows run SRGAN model

When I trained my own model using my database, it worked well. But unfortunately if I ran the code in windows anaconda，it showed the error like this:

It maybe some problem about pytorch multiprocessing in windows.
If someone knows how to solve it,pls help! Thx~

Missing keys in state_dict

I get the below issue while training with both PSNR and ESRGAN pre-trained models.

  RuntimeError: Error(s) in loading state_dict for RRDBNet:
 Missing key(s) in state_dict: "model.1.sub.16.bias", "model.1.sub.16.weight".

Is it because the pre-trained model is now not compatible with RRDBNet model?

Lower PSRN results

When I test the model on the Set5 dataset with the pretrained weights in RRDB_PSNR_x4.pth, I get a psnr score of 15.0664 as opposed to 25+ reported in the paper. Do you know what the issue might be? I haven't modified the training code and ran the image preprocessing scripts as reported in the paper.

The only difference is that I used octave instead of matlab to run the downsampling script. The script produced uint16 images, which I converted to uint8 using the following code:

    img = cv2.imread(v, cv2.IMREAD_UNCHANGED)
    img = img.astype(np.float32)
    img = (img * 255.0 / 65535.0).round()
    img = img.astype(np.uint8)
    cv2.imwrite(v, img)

Problem

I use your weight and ESRGAN model to test. I found the PSNR and SSIM is worse than that in the paper. Set 5 is 28.5438 PSNR, 0.8313 SSIM.

How can I modify the code to support 16bit .tif images?

I'm just wondering that is it possible for this model to be able to support 16bit .tif image? Or I have to convert it to 8bit?

Why is the input data randomly scaled in SFTGAN? Is it just for data augmentation?

Hi,

Just for clarification, why is the data randomly scaled in addition to the x4 upscale in SFTGAN? Is it just for purposes of data augmentation?

Also, is there a simple way of pre-generating the low-res images using mat-lab? The way the logic is implemented, (since it also randomly samples images from the DIV2k dataset), it seems impossible to simply put the LR images in a directory and complete the "dataroot_LR": data field in train_sftgan.json. Is that correct? I can understand ERGAN implementation but SFTGAN implementation is still a bit confusing to me.

Thanks

test_mode bug

I want to test the SR and SRGAN model. Now, I find a question :
Traceback (most recent call last):
File "test.py", line 49, in
for data in test_loader:
…
…
File "/opt/BasicSR-master/codes/data/util.py", line 262, in calculate_weights_indices
if not math.isclose(weights_zero_tmp[0], 0.0, rel_tol=1e-6):
TypeError: a float is required

I tried to tranform the weights_zero_tmp[0] to floattensor, but it still don't work.

weights_zero_tmp_1 = torch.sum((weights == 0), 0)
print(type(weights_zero_tmp_1)) #torch.ByteTensor
weights_zero_tmp = weights_zero_tmp_1.float()
print(type(weights_zero_tmp)) torch.FloatTensor
print(type(weights_zero_tmp[0])) torch.FloatTensor
print(weights_zero_tmp[0]) #120 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 120[torch.FloatTensor of size 18]#
if not math.isclose(weights_zero_tmp[0], 0.0, rel_tol=1e-6):
indices = indices.narrow(1, 1, P - 2)
weights = weights.narrow(1, 1, P - 2)
if not math.isclose(weights_zero_tmp[-1], 0.0, rel_tol=1e-6):
indices = indices.narrow(1, 0, P - 2)
weights = weights.narrow(1, 0, P - 2)

How can I continue learning from a checkpoint?

Did you implemented some mechanisms to continue learning when there was a failure?
If so, how can i use it?

thanks

Q: Real-time higher-res images (like a movie?)

Hi,

I was wondering is it possible with BasicSR/ESRGAN to create real-time 2x or 4x higher resolution from a video stream of 512x448 pixels with let's say 20 or 30 FPS?

Real-time, I mean a stream of input data, which creates a stream of output data (eg. like a movie as input).

Kind regards,
Melroy

The need of detach() on self.netF(self.var_H).detach()

Hi,

I've been reading your code and wonder if the detach function on NetF is necessary.
In SRRaGAN_model.py, line 128: real_fea = self.netF(self.var_H).detach(), since we don't need to update the NetF, is that OK to remove .detach() function?
If no, then why the fake_fea = self.netF(self.fake_H) doesn't have the detach function?

---

[Update]
Sorry but I think I understood your code. This line is for updating the NetG then we don't need to detach the self.netF(self.fake_H) (because self.fake_H is the result of NetG). Instead, we do self.netF(self.var_H).detach() since it is not necessary.
Please close this topic if my understanding is correct.

The create_lmdb.py doesn't work with the same filenames

Tried to create my own lmdb file using the create_lmdb.py file.
I found that the script only works with different filename cause it uses that filename (basename) as key.

My path is somewhat like this:

/home/Dataset/SISR/Class_A/XXXX.png
/home/Dataset/SISR/Class_B/XXXX.png
/home/Dataset/SISR/Class_C/XXXX.png

Filenames are the same but the content is different. The create_lmdb.py will create only 1 key (it means the size of dataset will be 1 instead of 3 files)

The only solution I have right now is to rename all image filenames so that they do not duplicate. Then run the code to generate the lmdb file. But it seems to be not the convenient way.
Do you have any better ideas? Thanks

---

[UPDATED] My solution is to add the iterator i with basename so that they will have a unique name.
I changed the line #33 in the create_lmdb.py to:
base_name = str(i) + "_" + os.path.splitext(os.path.basename(v))[0]

Do I need to modify the model to support scale x3 not x4?

Hi,

If I need scale x3 not x4, should I modify the model and train it from scratch? Where do I need to modify..?

Thank you

How to prepare train data for SFT-GAN?

I want to use DIV2K800 dataset to train SFT-GAN, but I don't know which folder should I put DIV2K800 segmentation .pth files? And does the OutdoorScene testing dataset require my manual classification?

About l_pix during training the sr model

When I use the your model to train the sr model. I found a problem, the l_pix is stable at around 0.02, is this normal? I have not changed any parameters, it is the same as your model.

The issue of compare_ssim function

@xinntao Hello, I found that you use the skimage.measure to compute SSIM index. But according to the offical API, this compare_ssim function should set gaussian_weight to True, sigma to 1.5, and use_sample_covariance to False to match the implementation of Matlab (Wang et. al.).
https://github.com/xinntao/BasicSR/blob/71635757505ef113bfba029441a4e35e59994a86/codes/utils/util.py#L90-L92

skimage.measure.compare_ssim

what about I change the architecture of the model?

If I change the architecture of the model, can I use the pretrained model RRDB_PSNR_x4.pth?

Confusion about l_d_cls_fake

I'm confused about the category loss "l_d_cls_fake = self.cri_ce(cls_d_fake, self.var_cat)", which seems the same behaviour as "l_g_cls = self.l_gan_w * self.cri_ce(cls_g_fake, self.var_cat)" in generator and violation of the nature of discriminator. Shouldn't it be "l_d_cls_fake = self.cri_ce(cls_d_fake, 'background')" or something? Thank you in advance.

error in training

Traceback (most recent call last):
File "C:/BasicSR/codes/train.py", line 158, in
main()
File "C:/BasicSR/codes/train.py", line 22, in main
opt = option.parse(parser.parse_args().opt, is_train=True)
File "C:\BasicSR\codes\options\options.py", line 18, in parse
opt = json.loads(json_str, object_pairs_hook=OrderedDict)
File "C:\Users\Mixtresh\AppData\Local\Programs\Python\Python36\lib\json__init
__.py", line 367, in loads
return cls(**kw).decode(s)
File "C:\Users\Mixtresh\AppData\Local\Programs\Python\Python36\lib\json\decode
r.py", line 339, in decode
obj, end = self.raw_decode(s, idx=_w(s, 0).end())
File "C:\Users\Mixtresh\AppData\Local\Programs\Python\Python36\lib\json\decode
r.py", line 357, in raw_decode
raise JSONDecodeError("Expecting value", s, err.value) from None
json.decoder.JSONDecodeError: Expecting value: line 1 column 1 (char 0)

i run train on windows 8.1

what is the dataroot_HR_bg in train_sftgan.json?

what is the dataroot_HR_bg in train_sftgan.json?

Local variable 'val_loader' referenced before assignment when training with train_sr.json

Hello,

I encounter the error UnboundLocalError: local variable 'val_loader' referenced before assignment when trying to train SR. The execution command is python train.py -opt options/train/train_sr.json.

Do any one have an idea why this happens?

The full output and error is as follows and is encountered after iteration 8 in the validation phase:

Path already exists. Rename it to [/home/.../.../basicsr-xinntao/experiments/debug_001_RRDB_PSNR_x4_DIV2K_archived_181127-104642]
Random Seed:  0
Dataset [LRHRDataset - Sobolt] is created.
Number of train images: 9,687, iters: 606
Total epoches needed: 1651 for iters 1,000,000
initialization method [kaiming]
---------- Model initialized ------------------
Number of parameters in G: 16,697,987
-----------------------------------------
Model [SRModel] is created.
---------- Start training -------------
/opt/anaconda3/lib/python3.6/site-packages/torch/nn/modules/upsampling.py:122: UserWarning: nn.Upsampling is deprecated. Use nn.functional.interpolate instead.
  warnings.warn("nn.Upsampling is deprecated. Use nn.functional.interpolate instead.")
<epoch:  0, iter:       2, time:1.31, lr:2.0e-04> l_pix: 4.21e-01
<epoch:  0, iter:       4, time:1.24, lr:2.0e-04> l_pix: 3.85e-01
<epoch:  0, iter:       6, time:1.24, lr:2.0e-04> l_pix: 3.41e-01 
<epoch:  0, iter:       8, time:1.25, lr:2.0e-04> l_pix: 2.01e-01 
Saving the model at the end of iter 8.
---------- validation -------------
Traceback (most recent call last):
  File "train.py", line 156, in <module>
    main()
  File "train.py", line 106, in main
    for val_data in val_loader:
UnboundLocalError: local variable 'val_loader' referenced before assignment

I am using 4 GPUs, a batch_size of 16 with 8 n_workers and an HR_size of 512.

Any help is appreciated.

Could you please upload your lmdb files to GDrive?

Hey Xinntao!
I'd like to work on your project and would like to ask you for your lmdb files so I can start faster.
Thank you very much!

About the Bicubic Interpolation

Thanks for your excellent repository! But I wonder if there is any difference between the OpenCV build-in Bicubic interpolation and the Bicubic algorithm you implemented in codes/data/util.py.

Where is the segmentation model OutdoorSceneSeg_bic.pth？

line 28 in codes/test_seg.py
load_path = '../experiments/pretrained_models/OutdoorSceneSeg_bic.pth'
Should we train a segmentation model by ourselves？

Could you explain a little bit about the statistic metrics?

This is a beginner question.. In the tensorboard, there are D_fake, D_real, l_d_fake, l_d_real, l_g_fea, l_g_gan, l_g_pix. Could you please explain what they are and what the expected trend each of them should have?

Thank you so much.

Issue with simple testing

I'm just getting into it and already got a problem with testing the pre-trained Esrgan model.

When running it I get the following error message which seems to be a windows related Pytorch issue:

File "C:\Users\Steven\Anaconda3\lib\multiprocessing\spawn.py", line 136, in 
_check_not_importing_main
is not going to be frozen to produce an executable.''')
  RuntimeError:
    An attempt has been made to start a new process before the
    current process has finished its bootstrapping phase.

    This probably means that you are not using fork to start your
    child processes and you have forgotten to use the proper idiom
    in the main module:

        if __name__ == '__main__':
            freeze_support()
            ...

    The "freeze_support()" line can be omitted if the program
    is not going to be frozen to produce an executable.

Any idea how to fix it?

Q：create lmdb Error：ZeroDivisionError: float division by zero

First time I ran the creat_lmdb, it occured that ZeroDivisionError: float division by zero.

if I change the code like that:

It can read, but fail to write. The error is the same... But!!! It produce the lmdb file even my cmd told me that it had error.So anyone help me please....

Problem with pth files

As I see, test_seg.py unsquzee's seg sensors to get 4D before saving. But train.py tries to read 3D tensors from pth files

Why texture changes frequently in GAN training?

Hi. Have you ever noticed that the detail or texture always changes frequently during training GAN? If yes, why???

About training segment model

Could you post the training set for segment model? It seems apart from COCO and ADE, you have combined some addtional images. And also, could you provide training details like hyper-parameters? I see Xiaoxiao's code but it looks like based on caffe. Since we have seg_arch.py, I'm not sure whether it is possible to train the segment model in Pytorch? Thanks in advance!

Invalid links of OST dataset

The links of OST cannot be opened.

do you have the tensorflow version of the code?

About esrgan Parameter

I saw some parameter in train_esrgan.json，but I am not sure the meaning. So pls can you give some detail to help me?

1.Like beta1_G,beta1_D,lr_gamma in train_esrgan.json?

2.I saw in the code have the training model for wgan-gp,so can it work? Because I saw ReadMe just delete this part. So I am not sure whether you change the code or not? Did you do the experiment to use the D as wgan-gp? If you did it before, could you tell me the perceptual quality use the D as wgan-gp?

3.I am a new in Neural Networks, and never try a model in image. So I want to know at least how many image can I build the esrgan model if I use 23 DDRB? And if I don't change the architecture, can I use my pre-train mode as pretrain_model_G,then change the parameter and run the code again to get a better network?

Please help，thank you so much...

Where is the option folder?

Sorry, I didn't find the options folder as well as config.json. Could you please check again ? Thank you

retrieve training procedure

if I have stopped training due to some technical reasons, and I have previous versions of my model saved to the 'experiments' folder, how can I start not from the beginning again but from where I stopped the last time?

Why to you detach fake_H?

Why do you detach the fake_H in SRRaGAN_model.py? The update would not be applied to G, as you did not add the G parameter to the optimizer. Is that correct?

pred_d_real = self.netD(self.var_ref)
pred_d_fake = self.netD(self.fake_H.detach()) # detach to avoid BP to G

optimizer:
self.optimizer_D = torch.optim.Adam(self.netD.parameters(), lr=train_opt['lr_D'], \ weight_decay=wd_D, betas=(train_opt['beta1_D'], 0.999))

ConnectionRefusedError: [Errno 111] Connection refused

when i run "train.py",it shows that "ConnectionRefusedError: [Errno 111] Connection refused"
line 156:main();then i add mp.set_start_method,but it not working

Rescaling image from 0..1 to -1..1

Is it possible to modify the code so the normalized pixels are in range -1..1?
I started to do it myself, but then stopped at

def rgb2ycbcr(img, only_y=True):
    '''same as matlab rgb2ycbcr
    only_y: only return Y channel
    Input:
        uint8, [0, 255]
        float, [0, 1]
    '''

won't it ruin the logic of the transformations?

The training time？

Hello, how much time do the training process spend? I find it slowly...

When I run the architecture.py alone, I get an error.

The following error pops up：

from . import block as B
ImportError: cannot import name 'block'

Thanks a lot!

spawn.py issues

when I run train.py I get:

19-01-02 21:35:49.755 - INFO: Model [SRRaGANModel] is created.
19-01-02 21:35:49.773 - INFO: Start training...
Traceback (most recent call last):
  File "train.py", line 154, in <module>
    main()
  File "train.py", line 79, in main
    for i, train_data in enumerate(train_loader):
  File "C:\Users\Claes\AppData\Local\Programs\Python\Python36\lib\site-packages\torch\utils\data\dataloader.py", line 819, in __iter__
    return _DataLoaderIter(self)
  File "C:\Users\Claes\AppData\Local\Programs\Python\Python36\lib\site-packages\torch\utils\data\dataloader.py", line 560, in __init__
    w.start()
  File "C:\Users\Claes\AppData\Local\Programs\Python\Python36\lib\multiprocessing\process.py", line 105, in start
    self._popen = self._Popen(self)
  File "C:\Users\Claes\AppData\Local\Programs\Python\Python36\lib\multiprocessing\context.py", line 223, in _Popen
    return _default_context.get_context().Process._Popen(process_obj)
  File "C:\Users\Claes\AppData\Local\Programs\Python\Python36\lib\multiprocessing\context.py", line 322, in _Popen
    return Popen(process_obj)
  File "C:\Users\Claes\AppData\Local\Programs\Python\Python36\lib\multiprocessing\popen_spawn_win32.py", line 65, in __init__
    reduction.dump(process_obj, to_child)
  File "C:\Users\Claes\AppData\Local\Programs\Python\Python36\lib\multiprocessing\reduction.py", line 60, in dump
    ForkingPickler(file, protocol).dump(obj)
TypeError: can't pickle Environment objects

E:\projects\BasicSR\codes>Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "C:\Users\Claes\AppData\Local\Programs\Python\Python36\lib\multiprocessing\spawn.py", line 105, in spawn_main
    exitcode = _main(fd)
  File "C:\Users\Claes\AppData\Local\Programs\Python\Python36\lib\multiprocessing\spawn.py", line 115, in _main
    self = reduction.pickle.load(from_parent)
EOFError: Ran out of input

I am running windows 10 home and python 3.6.8 gtx 1060 max q

Clarification for pretrained models

Are following statements correct?

• sft_net_ini.pth is a model that was trained for L1 loss only on DIV2K (+probability maps for each of the 800 DIV2K images)
• ImageNet/Flickr pictures were not used to training sft_net_ini.pth or any SFTGAN model.

activation for RRDB_net

the default activation_type you use in generator is 'leakyrelu'. have you tried doing as DCGAN paper suggests? they claim that one should use relu for all layers in generator except for the output where the tanh should be used.

I noticed that you have changed the project wiki

I noticed that you have changed the project wiki,could you please show the old project wiki about the data prepare and the options,thank you very much!!

Some questions about Sigmoid function

Hi, thanks for your novel work, I have some questions.

In the Relativistic GAN the sigmoid function is used little differently from standard GAN. Specifically, I see in the paper it says the Sigmoid function is added after C(X_r) - C(X_f). But I haven't found the Sigmoid function in your code. Could you please help me to find the location and usage of it?

Lower results

I used matlab to reduce the resolution of the image, but why the test results obtained with the pre-training model are two points lower than the results given.

pretrained model

I'm new to this. I'm curious that when I train the model, I got G.pth and D.pth. While in the option files, e.g., test_esrgan.json, you only provide option for "pretrain_model_G", then how about model_D? What if I want to use pretrained model_D?