lidq92 / wadiqam

[unofficial] Pytorch implementation of WaDIQaM in TIP2018, Bosse S. et al. (Deep neural networks for no-reference and full-reference image quality assessment)

(Wa)DIQaM-FR/NR

PyTorch 1.1 (with Python 3.6) implementation of the following paper:

Bosse S, Maniry D, Müller K R, et al. Deep neural networks for no-reference and full-reference image quality assessment. IEEE Transactions on Image Processing, 2018, 27(1): 206-219.

You can refer to the chainer codes (only the test part) from the original authors: dmaniry/deepIQA

Note

• The hyper-parameter or some other experimental settings are not the same as the paper described, e.g., nonoverlapping patches are considered for validation/test images instead of random selection. Readers can refer to the paper for the exact settings of the original paper.
• Warning!. The performance on each database is not guaranteed using the default settings of the code. Reproduced results are welcomed to reported.
• If you do not have enough memory, then change slightly the code in IQADataset class. Specifically, read image in __getitem__ instead of __init__. You can choose to use IQADataset_less_memory class instead.

TODO (If I have free time)

• Reproduce the results on some common databases, especially for the NR model (Currently, NR model is not tuned to reproduce the results.)
• Simplify the code
• etc.

wadiqam's Issues

Could you please provide a pretrained model for No-reference IQA?

Hello, thanks for sharing your good work. I only found full-reference model in checkpoint dir. Could you please provide a pretrained model for No-reference IQA?

您好，感谢您提供的代码~~~这儿我有个小问题

在您的test_cross_dataset.py这个测试脚本中最后有个循环
for i in range(len(im_names)):
im = Image.open(os.path.join(args.dist_dir, im_names[i])).convert('RGB')
ref = Image.open(os.path.join(args.ref_dir, ref_names[i])).convert('RGB')
这个循环是不是只能测试25张有参考的iqa？上述语句中的i是同时在变的，也就是说应该没法儿测试比如第1张参考图像对应第4张失真图像的iqa分数？想跟您确认一下这个点，感谢

Inquire about the data info generation process

Dear Dingquan,

Thanks a lot for your great work! It really helps!

One thing I wanna inquire about is the data info generation process. Is any script available to generate such full info for a random dataset like CSIQ?

Thanks in advance and have a nice day!

Best,

Shuyue

License info

Hello is the pre-trained model free to use for commercial use as part of a transfer learning model?

result higher than the paper

hi！I would like to ask you about the replication results of using this code on the Tid2013 dataset. The results I ran were much higher than those in the paper. What is the situation? May i ask what data you have generated?

Query regarding DIQaM

Thanks for your work in this field of IQA. Can you please tell me what changes do I need to bring in this code if I want to implement DIQaM method of the paper?

Failed to save model graph

def create_summary_writer(model, data_loader, log_dir='tensorboard_logs'):
    """ Create Summary Writer for TensorBoard Visualization
    :param model: model
    :param data_loader: data loader
    :param log_dir: log dir
    :return: writer
    """
    writer = SummaryWriter(log_dir=log_dir)
    data_loader_iter = iter(data_loader)
    x, y = next(data_loader_iter)
    try:
        writer.add_graph(model, x) # exception here
    except Exception as e:
        print("Failed to save model graph: {}".format(e))
    return writer

Error:
Failed to save model graph: %33 : Tensor = onnx::Unsqueezeaxes=[0]
has empty scope name

Could you help me ? my version is torch 1.1.0 torchfile 0.1.0 torchnet 0.0.4 torchvision 0.3.0

您好，我在您的网络基础上改动为NRIQA遇到了一些问题

您的FRIQA网络我已经成功训练并测试，之后我想要将这个网络修改为无参考的模型进行训练并测试，我做了如下的改动：将配置文件的模式修改为WaDIQAM-NR，将RandomCropPatches切片函数的参数修改为NR，然后运行训练脚本，切分数据集的过程中会产生如下问题：RuntimeError: expected a non-empty list of Tensors，我尝试着将RandomCropPatches函数的ref_crops删除，还有IQADataset中的有关ref_crops的部分都做了删除，并将model中的x_ref删除，只输出其他的值，之前的问题得以解决，但是会出现这个问题：
Traceback (most recent call last):
File "D:/pythonworkplace/MY-WAdeepIQA/WaDIQaM.py", line 138, in
sq_std.append(data[3])
IndexError: list index out of range
面对这个问题我多次尝试但是没有能够成功解决，如果将[3]删除虽然能解决掉字面的bug但是在运行一段时间后还会提示只有一个tensor元素可以被转化为python scalar的错误。究其原因还是我对于代码中定义的sq_std的意义不明确，不知道在无参考模型中如何处理这个变量，希望能够得到您的指点，帮助我完成无参考的训练，万分感激！！