[ICCV 2023] Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution; runner-up method for the model complexity track in NTIRE2023 Efficient SR challenge
大佬您好，请问在哪里可以测量参数、激活数等的这些指标？我在您的源码中没有找到关于测试这个指标的代码
另外这3个模型有什么区别（我是程序员，想用这个超分辨率，不太懂算法）
BI-Efficient SR
BI-Classic SR
x4 High-order
Hi, authors
Thanks for your great work! The Fig.2 of LAM visualiztion is a very impressive way to demonstrate the different receptive filed between different models. And I also want to implement this, so i wonder is it convenient for you to share the LAM visualiztion code? It will really helps a lot to me :D
Hello, I would like to ask a question:
How is the inverted residual block module defined in ablation experiments? Because I remember that in MobileNet v2, this module has a very small number of parameters.But I saw that in your article's ablation experiment, replacing CCM with inverted residual block significantly increased the parameter quantity.
I would greatly appreciate it if you could help clarify my confusion. Thank you!
I have already loaded the data and pre training model, and I encountered the following error while running the code：
Name RealSRModel_basicsr is not found, use name: RealSRModel_basicsr_basicsr!
Traceback (most recent call last):
File "basicsr/test.py", line 45, in
test_pipeline(root_path)
File "basicsr/test.py", line 35, in test_pipeline
model = build_model(opt)
File "/opt/conda/lib/python3.8/site-packages/basicsr/models/init.py", line 26, in build_model
model = MODEL_REGISTRY.get(opt['model_type'])(opt)
File "/opt/conda/lib/python3.8/site-packages/basicsr/utils/registry.py", line 71, in get
raise KeyError(f"No object named '{name}' found in '{self._name}' registry!")
KeyError: "No object named 'RealSRModel_basicsr' found in 'model' registry!"
在RealGANModel模式下，各个损失的loss_weight应该如何设置比较合理
ref chaiNNer-org/chaiNNer#2504
在上述项目遇到的问题。
How can I download test sets such as set5？
您好，在使用网络时碰见这个问题：第一次完整使用网络训练：python basicsr/train.py -opt options/train/SAFMN/train_DF2K_X2.yml
后未发现异常，可以正常生成结果。但是想重新训练时，出现如下报错：AssertionError: An object named 'SAFMN' was already registered in 'arch' registry!。在train_DF2K_X2.YML中network_g->type修改为SAFMN_NITRE后报错信息仍不变，请问如何解决这个问题，感谢！
大佬您好，跟您请教一下，您的算法是在RGB图像直接处理的还是转化到YCbCr上再进行处理
Trying to test for a single image I have in SAFMN\datasets\real_test\image.jpg. Not sure how the parser works so I'm struggling to find why this error is thrown.
inference code for this?
您好，我想请教一些问题：
1、消融实验中SAFM模块是怎么验证的？ 去掉的意思是完全不用这个模块还是 换成别的模块替代呢
2、您的特征图可视化是怎么做的呢？
非常希望您能解答我的疑惑，谢谢！
孙博士您好，我有个问题需要您指导，您在 options/train/SAFMN /train_DF2K_X2.yml中在验证集中的设置为datasets/DIV2K/DIV2K_Valid/LR_bicubic/X2，我理解的bicubic可以通过matlab或者opencv操作，请问后面的X2要放什么内容，是和训练集一样把图片分辨率降低后的图片放到这个文件夹下吗？
还有一个问题是在进行测试时，train_DF2K_x2.yml文件夹内：
datasets:
test_1: # the 1st test dataset
name: Set5
type: PairedImageDataset
dataroot_gt: datasets/Benchmarks/Set5/HR
dataroot_lq: datasets/Benchmarks/Set5/LR_bicubic/X2
filename_tmpl: '{}x2'
datasets/Benchmarks/Set5/LR_bicubic/X2存放的内容是什么？
谢谢您的指导！
大佬您好，打扰您一下，您是在DIV2K和Flickr2K数据集上训练的，如果只想在DIV2K上训练的话，请问要往哪方面调参数效果会比较好
您好！
我用 Nvidia T4 (性能接近2080Ti) 做了一个推理测试，测试程序是 sample代码app.py, 图片大小2560*1440,scale为2倍模型，处理性能大概是9s多，是不是哪里处理的有问题？或者如何优化？
谢谢！
作者您好，请问您real-world模型还是用同样参数量的小模型，follow real-esrgan的方法训练的吗？还是有给模型参数量加大
您好，我在Google Colab上成功运行了4被超分和2倍超分模型。然后我用
if test_patch_wise:
!python inference/inference_real_safmn.py --model_path experiments/pretrained_models/SAFMN_DF2K_x3.pth --input BSRGAN/testsets/RealSRSet --output results/SAFMN_x3 --scale 3 --large_input
else:
!python inference/inference_real_safmn.py --model_path experiments/pretrained_models/SAFMN_DF2K_x3.pth --input BSRGAN/testsets/RealSRSet --output results/SAFMN_x3 --scale 3
试图运行3倍超分，结果得到了“预训练模型参数不匹配”的问题。请问应该如何解决，谢谢
您好，打扰您一下，找不到您比赛时完整的预训练模型（x2 , x3 , x4），在您的文件里只找到了一个关于您的team15_SAFMN.pth文件
hi, I test SAFMN_L_Real_LSDIR_x4.pth on my own picture， but some pic color is changed， how can I solve this problem?
您好！我在训练时设置batch_size_per_gpu=8训练时间为16h，batch_size_per_gpu=48训练时间为16d，两者CPU均为满负载，是因为CPU IO瓶颈了嘛？将数据集读取格式为lmdb也没有影响。请问是哪里出了问题？
您好，我使用set5与set14作为验证集，但是在同一model上验证与测试的psnr与ssim不一致，请问这是为什么？
Hi, thank you for your hard work!
I am trying to convert SAFMN to onnx but there is a problem with adaptive_max_pool2d
class SAFM(Module):
def __init__(self, dim, n_levels=4):
super().__init__()
self.n_levels = n_levels
chunk_dim = dim // n_levels
# Spatial Weighting
self.mfr = ModuleList([Conv2d(chunk_dim, chunk_dim, 3, 1, 1, groups=chunk_dim) for i in range(self.n_levels)])
# # Feature Aggregation
self.aggr = Conv2d(dim, dim, 1, 1, 0)
# Activation
self.act = GELU()
def forward(self, x):
h, w = x.size()[-2:]
xc = x.chunk(self.n_levels, dim=1)
out = []
for i in range(self.n_levels):
if i > 0:
p_size = (h//2**i, w//2**i)
s = torch_nn_adaptive_max_pool2d(xc[i], p_size)
s = self.mfr[i](s)
s = torch_nn_interpolate(s, size=(h, w), mode='nearest')
else:
s = self.mfr[i](xc[i])
out.append(s)
out = self.aggr(torch_cat(out, dim=1))
out = self.act(out) * x
return out
I even tried replacing adaptive_max_pool2d with max_pool2d but it still gives an error, is there a way to solve this? :)
Hi, thanks for publishing the code of your work! I am running the test code using commands in your README.md and facing some issues. I used the scripts inside the "scripts" folder to download pre-trained weights and the Set5 dataset.
My current confusion is: Which basicsr module to use? README asks to install it from pip but there is also one packaged inside the repo. Of course, python uses the one from pip by default but is the packaged one different (possibly causing unexpected bugs)?
大佬你好， 我在a100上跑你们的网络模型要花3天多的时间，这个正常吗？
根据提供的training log，batch_size_per_gpu=64（lightweight SR）和batch_size_per_gpu=48（classic SR）为什么训练需要的epoch一样啊？
Excellent work, I can't wait to try training this network with my data, looking forward to your release of the source code
对于数据集B100 Urban100 Manga109，只能下载到HR，我怎么才能获得LR_bicubic呢
万分感谢
FileNotFoundError: [Errno 2] No such file or directory: 'experiments/pretrained_model/SAFMN_DF2K_x4.pth'
因为服务器连不了外网，麻烦告诉我去哪里下载预训练模型，
作者您好！
看到您在文中用了FFT loss，我训练了两三万iter简单试了一下，发现删掉FFT loss指标好像会有所提高，想问问您有没有验证过fft loss对完整训练流程的影响，删去fft loss会导致模型最终性能的下降吗？
作者你好，想请问下，你前面issue里说了real的yml效果好一些，但现在的branch里没有这个配置，只有test有，train没有；
另外，你这些train_DF2K_X4.yml， 还有下划线_L 标识的，和classic，和real是什么意思，模型都是用的SAFMN吗，只有dim深度不同是吗？
我想用最好的是选哪个，怎么会配置？
想用论文或者比赛里的，怎么配置？
想用超轻量的，怎么配置？
Hi, how do I prepare datasets for training. Say that I want to train the efficient model using DIV2k dataset? What dataset did you use to train the SAFMN_L_Realx4.pth -model?
请问作者在训练的时候有没有出现损失变为nan的情况呢？如果有的话是怎样解决的呢？
Hello, I am really interested in your research. I find that there are three different models in your NTIRE2023_ESR. Could you tell me which one is the best model? Thank you very much!
May I ask how to solve the title problem
你好，不单独安装nvdia驱动，使用自带的nouveau作为gpu驱动，可以跑出来结果吗
whhich model do you upload to OpenXlab? can you share the training details?
Fantastic project but out of memorry for bigger pics.
cant upscale over 3MPX with 12GB GPU Ram
Thank you very much in advance
大佬，想问下：在set5数据集上，X4SR，平均推理时间达到了0.5s左右，这不正常吧。会是什么原因导致？
Hi there, thanks for your great work!
I noticed your pretrained models had a 'SAFMN L real' one, how to train this model? because i want to train a 'x2 efficient real' model myself.
您好，
感谢您的精彩论文，既有很好的创新，也很实用。我想用这个网络尝试跑这个数据集（NTIRE 2023 Efficient SR challenge），请问如何设置训练和测试的YML文件呢？
您好！对您提供的SAFMN_Realx2.pth在真实图像中的超分表现我们感觉效果很好，但是这个在Set5数据集上的评估PSNR就比较低，
2024-04-20 17:37:07,104 INFO: Loading SAFMN model from D:/pythonSoftware/codes/SAFMN/models_pretrain/SAFMN_L_Real_LSDIR_x2.pth, with param key: [params].
2024-04-20 17:37:07,190 INFO: Model [SRModel] is created.
2024-04-20 17:37:07,191 INFO: Testing Set5...
2024-04-20 17:37:08,370 INFO: Validation Set5
# psnr: 25.5341 Best: 25.5341 @ SAFMN_c36n8_x2 iter
# ssim: 0.8602 Best: 0.8602 @ SAFMN_c36n8_x2 iter
所以i想请教您SAFMN_Real是如何训练的呢？ 谢谢！！！
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
Django
The Web framework for perfectionists with deadlines.
Laravel
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
Microsoft
Open source projects and samples from Microsoft.
Google
Google ❤️ Open Source for everyone.
Alibaba
Alibaba Open Source for everyone
D3
Data-Driven Documents codes.
Tencent
China tencent open source team.