Why is the code different from the loss function in the paper? Q_loss is completely different from LQ in the text.

Hello, because I can't download the pretrained model in China, can you will advance training model to my E-mail: [email protected]
Thank you very much!

hi,

The paper doesn't use any pairwise loss. Is this a joke? why are you implementing something different that the paper's claim?

thanks in advance

Thanks for the great work!
I 'm interested in your video hashing work. Can you release your video hashing code?
My email is [email protected]. Hope to communicate with me.

hi~ I want to train CSQ on a person reid task.
the class num is more than 40,000
the code which was used to set the hash target center can be extremly time consumming,
do you have any suggestion?
@yuanli2333

`

  if H_2K.shape[0] < n_class:
        hash_targets.resize_(n_class, bit)
        for k in range(20):
            for index in range(H_2K.shape[0], n_class):
                ones = torch.ones(bit)
                # Bernouli distribution
                sa = random.sample(list(range(bit)), bit // 2)
                ones[sa] = -1
                hash_targets[index] = ones
            # to find average/min  pairwise distance
            c = []
            for i in range(n_class):
                for j in range(n_class):
                    if i < j:
                        TF = sum(hash_targets[i] != hash_targets[j])
                        c.append(TF)
            c = np.array(c)

            # choose min(c) in the range of K/4 to K/3
            # see in https://github.com/yuanli2333/Hadamard-Matrix-for-hashing/issues/1
            # but it is hard when bit is  small
            if c.min() > bit / 4 and c.mean() >= bit / 2:
                print(c.min(), c.mean())
                break

`

I have a question that how ro create the train.txt and test.txt the format of which are different from my own dataset . And what's the meaning of the hyper-parameters: "lambda0","lambda1","lambda2". Thank you!

Hello, how can I solve this problem?

I run the source code,the train.py and test.py,and the results are great.And the paper show a example which give an image to find other similar images, but it seems the code doesn't include it. Could you tell me how to achive it?Is it just take an image as input or need label?

Hi, thank you for good work.

I have some question on ImageNet model training.
I think the model use ImageNet pretrain weight for ImageNet training. [The code line]

If there are options for training ImageNet which I missing, or wrong information, please let me know.

Anyway, Thank you for your wonderful work agin.

hello!
Thanks for the great work!
I 'm interested in your hashing work. I tried to run this code with AlexNet for cifar10，but it didn't work well.
Can you release your parameters for cifar10 or give some advice for running in cifar10 with AlexNet ?

Hi,

First and foremost, amazing work! Thanks for having the code and models publicly available.

I was wondering if you could have the dataset/Kinetics directory and train_kinetics.py file uploaded as well?

Regards
Arun George

The crop size is set to be 224x224, which is not the same crop size as in HashNet (227x227). Is there a reason for that?

Why is the code different from the loss function in the paper? Q_loss is completely different from LQ in the text.

Most deep hashing methods adopt AlexNet as the backbone. For a fair comparison, can you share the results of this method under AlexNet?

May I know your specific parameters in NUS_ Wide？I use your library's parameters to run out the result is only 0.8188 using resnet model and 64 bit hash code in NUS_ Wide dataset

Why is the database.txt file of ImageNet not given? Can you give me a database.txt file?

hello, I‘am in China. I can't download pre-trained models by links your provided. Could you provide other download links for pre-trained models? Thank you.

Hi,I'm wondering why ablation study in paper doesn't give result of using L_Q loss only.If the result is obviously worse than using L_C only,it will demonstrate what a role that L_C is playing in this task.
Also,I'm also curious about why using a loss that forcely pull prediction points to predefined hash center,which is a singular point that is hard to converge to(and I know that part is talked in paper).I just want to know why create a such a loss function and how much the importance is.
Your reply will be highly appreciated.

Thanks for the great work!

Super excited to play around with video hashing. Are there any plans for releasing the video code and configurations?

can please share the train parameter for Alexnet on coco?

thanks for you work. I test pair image by the model:imagenet_64bit_0.8734_resnet50.pkl. and get the hamming distance:0
This is my test code:

class CSQ(object):
    def __init__(self):
        self.model_path = 'checkpoint/imagenet_64bit_0.8734_resnet50.pkl'
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        self.transform = prep.image_test(resize_size=255, crop_size=224)
        self.load_model()

    def load_model(self, ):
        self.model = torch.load(self.model_path)
        self.model = self.model.module
        self.model.to(device=self.device)
        self.model.eval()

    def forward(self, img_path):
        tensor_img = self.transform(Image.open(img_path).convert('RGB')).unsqueeze(0).to(self.device)
        with torch.no_grad():
            out = self.model(tensor_img)

        hash_code = out.cpu().numpy()
        hash_code[hash_code < 0] = -1
        hash_code[hash_code >= 0] = 1

        code_list = ''.join(['1' if item == 1.0 else '0' for item in hash_code[0].tolist()])
        return code_list

Hi,

You set the value min(c)=20 for ImageNet in '.ipynb' file.
How to set the value of min(c) for other datasets, such as COCO or NUS_WIDE?

How can I get my own 'database. txt', 'train. txt' and 'test. txt' when training my own dataset? What is the relationship between them? For example, are 'train. txt' and 'test. txt' included in 'database. txt'? Then what do the 0 and 1 after the file name in these three files represent?

I really need your answer，Thank you very much!

Hi, first really appreciate for sharing codes. I have a question about the experimental settings.
in your paper, experimental settings for imagenet is

ImageNet image 10,000 5,000 128,495 100:1

where the number of training images is 10,000. However, the number of training images in this repo is 13,000.
Which one is right to reproduce your result in the paper?

Hi~, I am a student. I am studying the video hash. I have read the article . the model of yours performs well. But I have trouble in reimplementing the other models. Could you provide codes of other models. Really appreciate for sharing codes. My email is :[email protected]

Thank you so much!