zwx8981 / dbcnn-pytorch Goto Github PK

我也在学习图像质量评价，我在知网上搜到了你的博士论文《基于特征聚合和数据驱动的盲图像质量评价》，但是知网上下载需要60元手续费不知道能否您私发我以下，哈哈，苦逼学生就是这么抠搜
方便的话，发我QQ邮箱可以吗？ [email protected]

您好，我使用您的代码在pytorch环境下运行DBCNN.py进行学习时，有一个报错
y = torch.tensor(y.cuda(async=True))
^
SyntaxError: invalid syntax
具体位置在DBCNN.py的第204及246行，期待得到您的回答！！

您好，看到您的更新。目前应该是把Waterloo、VOC的合成失真数据，换成了KADIS-700k合成失真数据集。能否提供数据集的下载链接，以及基于此数据集的预训练S-CNN模型。
非常感谢~

why do I get this error when I run your code: D:\zwx_Project\zwx_IQA\dataset\databaserelease2\jp2k\img12.bmp" does not exist , where is this address in the code , I couldn't find it

Hello author, I want to train scnn.py. But the model I needed for training is not there. How do I get this model?

Hi,
Are you going to make the pre-trained weights used in the paper available for DBCNN?

Hello, thanks for your code !
Can you tell me the accuracy of SCNN's classification of 39 distortion types? If I want to retrain SCNN, what is the appropriate accuracy( I can stop my training when I get it)?

Hi, me again.
Can you suggest/give more detail how to train on LIVE Challenge to get SRCC around 0.85?

I've trained on LIVE release2 using your code and got SRCC very close to 0.968 but on LIVE Challenge, I could get just around 0.81. I've tried changing learning rate, number of epochs, and using SGD instead of Adam on both FC train and FULL train. Did I miss anything or anything I should try?

Thank you in advance, really appreciate it.

Hi author:

I want to use simple_demo.py to test jpg quality. I set:

scnn_root = 'pretrained_scnn/scnn.pkl'

// ckpt = "your path of the checkpoint file"
image_name = "4.jpg"
// checkpoint = torch.load(ckpt)
// model.load_state_dict(checkpoint)

I comment the code of checkpoint, since it has been set in scnn_root, am I right?

“4.jpg” is my test image, and I get the output: Prediction = -0.0054. I think the prediction should be [0, 1], so whats wrong with my codes?

Kang

if i want to train dbcnn on koniq-10k, which code should i run? BaseCNN.py or DualCNN.py?

您好，我有几个问题请教：
1、LIVE数据集crop_size=432，那么训练出来的模型，只是针对432432的图片而言的。但是在计算PLCC和SROCC的时候，也是432432，那么这个PLCC只是当LIVE数据集crop_size=432时的PLCC、SROCC而不是原图的PLCC、SROCC。如果要得到真正LIVE数据集上的PLCC，需要输入原图获得预测质量，再去计算。是这样吗？
2、如1所述，LIVE的数据集训练出来的模型是针对432*432的图像输入。那么用这个模型，去预测任意尺寸的图像质量，这里有没有一个合理的说法？
3、可否提供一版自建数据集的训练代码
感谢分享，会持续关注您的工作~

Thank you so much for you code.

I'm curious that in Class LIVEFolder, you refer to 'dmos_realigned.mat' under Live dataset databaserelease2 but I found only 'dmos.mat' under the folder. Is it the same file or is there any additional process?
Thank you.

你好，最近我根据您论文中的思路和github代码进行了复现。目前想尝试修改SCNN架构，但是缺少对应的数据集，查阅了相关的代码https://github.com/zwx8981/DBCNN，由于能力有限，没看懂如何生成训练SCNN预训练模型的数据集（包括失真图片和label），您能否提供python版本生成预训练数据集的代码，不胜感激。
还有一个问题，对于DBCNN的优化，您是否有一些思路。

您好，我想请问一下关于代码运行问题。我如果想在LIVE数据库上训练的话，是应该先利用MATLAB的数据生成函数生成数据类型然后训练SCNN，还是可以直接训练DBCNN呢

Can you provide the classification accuracy of SCNN, thank you！

such as torch.Size([1, 3, 764, 960])
got
File "DBCNN.py", line 89, in forward X2 = X2.view(N, 128, H*W) RuntimeError: shape '[1, 128, 2820]' is invalid for input of size 368640
is that means i just can use square image to test？

文章里面的cross database部分写的是"All experiments are conduct by training models on one entir database and test them on th other databases"，那么如何做呢？意思是在live上分80%训练，20%测试的model直接在整个csiq，tid2013或者livec上面测试吗？还是应该用其他方法呢？

how could i solve it ？

Good work,Thanks for sharing!!! And I have some questiones:
1 、Have you tried some lightweight net in ImageNet replacing VGG-16? such as mobilenet、shufflenet. And how about the result comparing with vgg-16
2、what the function of the "net_params_best.pkl" in content "fc_model"
3、Do you plan to add distortion kind,such as "sharpness",in order to assess whether the enhance image is good or bad
Again,So excellent you work is!!!
You can repay with Chinese if possible.

I fine-tuned the DBCNN on LIVE dataset. ( python DBCNN.py and I didn't change the code.)
I used the models saved in 'db_models' folder, ( net_params_best.pkl ) to test on waterloo.
The results are as follows:

There is a gap with your accuracy given in paper, What's the possible reason for this, what should I do to improve it.

加载提供的预训练模型测试时，出现预训练模型参数和模型不匹配，可以加个联系方式吗

Hi. I have a custom dataset consisting of 5248 images. I made necessary changes to include this in data loader.

In dbcnn.py, I provide the list index as (0,5248). Dbcnn.py runs and takes several hours but does not produce even a single epoch.

When I change the list index as (0,29), then dbcnn.py produces training and val images and the script runs and produces 50 epochs that is set

The images that I use are taken from Android or iphone. Each image have different resolution. Can. You suggest a way to train on my images only? Is there any other way I can increase the speed of training by pre processing my images for your model?

论文的table 2，table 3和table 4的individual distortion types是怎么做的实验啊？比如说像table 2测live数据集的时候，是不是先把model在整个live上进行和table 1一样的分80%和20%的训练测试，然后保存model后对这5个失真类型的所有图片进行SROCC的打分呢？这部不是很清楚，请教一下，谢谢！

当运行DBCNN.py时报错：
Traceback (most recent call last):
File "/home/dell/LGL/DBCNN-PyTorch-master/DBCNN.py", line 354, in
main()
File "/home/dell/LGL/DBCNN-PyTorch-master/DBCNN.py", line 336, in main
manager = DBCNNManager(options, path)
File "/home/dell/LGL/DBCNN-PyTorch-master/DBCNN.py", line 169, in init
transform=train_transforms)
File "/home/dell/LGL/DBCNN-PyTorch-master/LIVEFolder.py", line 61, in init
self.labels = self.dmos['dmos_new'].astype(np.float32)
KeyError: 'dmos_new'

请问这是哪里出了问题，万分感谢！

I modified the DBCNN.py script and the data loaders to this such that it is trained for my images. It runs for 50 epochs and throws me an error - out of memory. Buy more ram.

The DBCNN.py runs for 50 epochs

Q1. How many epochs have you fixed?
Q2. Once 50 epochs is done, the model is saved under fc_models/net_params_best.pkl

When I use this .pkl file in the simple_demo.py, I get errors in state_dict - Missing keys in state_dict for DataParallel. How do I proceed with this?

Hello! We kindly invite you to participate in our video quality metrics benchmark. You can submit DBCNN-PyTorch to the benchmark, following the submission steps, described here. The dataset distortions refer to compression artifacts on professional and user-generated content. The full dataset is used to measure methods overall performance, so we do not share it to avoid overfitting. Nevertheless, we provided the open part of it (around 1,000 videos) within our paper "Video compression dataset and benchmark of learning-based video-quality metrics", accepted to NeurIPS 2022. It is available through the link (password: c943=R3/tJwVV%P%).

Hi, I am performing inference for my dataset using your model. What is the output range for your model? Is it [1-10] or [1-100]?

Good work,Thanks for sharing!!!
I was wondering if there are any available trained models on Live IQA dataset. I would appreciate if you share these ones for public access.

Kindly,
SManna

Hi author:
I would like to make use on the pretrained SCNN for the further training.
However, I found out the pretrained 'scnn.pkl' file in this repo is only a string of numbers instead of model weights.
Where can I find the Pytorch weights for SCNN, or will this be publically available?

I had run successfully model and got FC model DB model. How can we utilized these models to predict score?

when I uesed Gblur part in the LIVE database to train the network, I got only 0.3 srcc at most on its 20% test. I guess the reason is lack of data, so I split the picture into 100*100 patch, also get low srcc, could you give me some advice~, thank you

您好，我的数据集中不包含失真类型，只有图像和相应的mos值，请问我能进行训练吗