您好！　请问FLICKR-25K.mat数据可以分享下吗，我是新手，或者可以教下我怎么处理吗

Hi,
Thanks for your code.
I have a question that the loss is NAN from 105 epoch in the data set FLICKR-25K.
Can you have any idea about the problem?
I use the tensorflow code.

Many thanks!

ref:
...epoch: 102, loss: 27617250.477, comment: update B
...epoch: 103, loss: 27464020.237, comment: update B
...epoch: 104, loss: 27408385.730, comment: update B
...epoch: 105, loss: nan, comment: update B
...epoch: 106, loss: nan, comment: update B
...epoch: 107, loss: nan, comment: update B
...epoch: 108, loss: nan, comment: update B

Hello, I notice that you annotated some code in the function '' train_img_net ", I wonder if the results of the two strategies for index construction are the same? I hope you can help me understand it. Thank you.

# index = range(0, num_train - 1, 1)
for iter in xrange(num_train / batch_size):
	index = np.random.permutation(num_train)
	ind = index[0: batch_size]
	# ind = index[iter * batch_size: (iter + 1) * batch_size]
	unupdated_ind = np.setdiff1d(range(num_train), ind)

I found that the model fails to re-produce the results when using the commented-out sampling strategy in Ln 37 and Ln 41 in DCMH_demo.py, i.e., sampling WITHOUT replacement, even if I shuffle the index before each training epochs. And it works well with Ln 39 ~ 40, i.e., sampling WITH replacement.
But why ?

hello, I'm asking a details of the partition of the NUS-WIDE data set.
before that, did you follow the settings like randomly select XXX samples in each classes for query set? if yes, then my question is:
how do you randomly select 100 samples in each classes for the query set (and so does the training set)?
I mean, as it's a multi-label data set, a same very sample might be selected several times during the sampling of each classes.
thanks

Hi

I am trying to run the code for bit = 64 you provided in the tensor flow environment. The loss is decreasing but the accuracy (MAP) is remaining the same. As you can observe below the loss has decreased quite significantly but the accuracy is renaming the same.

Can you please tell me if I am going wrong somewhere ?

...loading and splitting data finish
...epoch:   0, loss: 177061069.097
...training procedure starts
...epoch:   1, loss: 175336058.349, comment: update B
...training procedure finish
...test map: map(i->t): 0.543, map(t->i): 0.543
...epoch:   2, loss: 173065982.115, comment: update B
...training procedure finish
...test map: map(i->t): 0.543, map(t->i): 0.543
...epoch:   3, loss: 170820578.049, comment: update B
...training procedure finish
...test map: map(i->t): 0.543, map(t->i): 0.543
...epoch:   4, loss: 168648051.304, comment: update B
...training procedure finish
...test map: map(i->t): 0.543, map(t->i): 0.543
...
...epoch: 496, loss: 70586304.761, comment: update B
...training procedure finish
...test map: map(i->t): 0.543, map(t->i): 0.543
...epoch: 497, loss: 70586286.625, comment: update B
...training procedure finish
...test map: map(i->t): 0.543, map(t->i): 0.543
...epoch: 498, loss: 70586388.189, comment: update B
...training procedure finish
...test map: map(i->t): 0.543, map(t->i): 0.543
...epoch: 499, loss: 70586918.393, comment: update B
...training procedure finish
...test map: map(i->t): 0.543, map(t->i): 0.543
...epoch: 500, loss: 70586985.325, comment: update B
...training procedure finish
...test map: map(i->t): 0.543, map(t->i): 0.543

In addition I had to make the following changes to run your code in Tensor-flow :
[1] changed xrange to range as I am using python 3
[2] put print statements inside quotes
[3] changed statements such as this
for iter in xrange(num_train / batch_size):
to
for iter in range(np.int(np.floor(num_train / batch_size))):
to make the data type int
[4] defined the variable
L['retrieval'] = labels[QUERY_SIZE:DATABASE_SIZE + QUERY_SIZE,:]

when i use vscode run the code of tensorflow,The following problems arise。But if i use pycharm or run it in DOS, it don't arise. so i can't find the reason of this BUG.Have you ever encountered this problem
problem：
Traceback (most recent call last):
File "main1.py", line 150, in
images, tags, labels = loading_data(DATA_DIR)
File "/content/drive/APP/load_data.py", line 7, in loading_data
image = file['images'][:].transpose(0,3,2,1)
File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper
File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper
File "/usr/local/lib/python3.6/dist-packages/h5py/_hl/group.py", line 177, in getitem
oid = h5o.open(self.id, self._e(name), lapl=self._lapl)
File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper
File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper
File "h5py/h5o.pyx", line 190, in h5py.h5o.open
KeyError: "Unable to open object (object 'images' doesn't exist)"

Hi, I wonder the classes order of the label vector and their corresponding class name of the NUS-WIDE data provided in DCMH-CVPR2017/DCMH_tensorflow/DCMH_tensorflow/readme.txt.
Or can you provided the processing code of NUS-WIDE ?
thanks

Hello, thanks for the code.
I have a question about the learning rate of SGD in the NUS-WIDE dataset. I am trying to use your code in the NUS-WIDE dataset, but I found that the result is so bad when using the same SGD optimizer and learning rate as the MIRFLICKR-25K dataset.

Hi! I want to know how the original image from flickr25k convert to 3224224 matrix, because I have want to convert matrix to image.

您好，预训练网络vgg_net.mat的下载链接好像失效了

Hello，thanks for the code.
I have a question about the learning rate of SGD in the NUS-WIDE dataset. I am trying to use your code in the NUS-WIDE dataset, but I found that the result is so bad when using the same SGD optimizer and learning rate as the MIRFLICKR-25K dataset.

你提供了MIRFLICKR25K、NUS-WIDE的下载地址，并说明了将多个文件合并为一个文件，
请问问合并文件的方法可以提供吗？或者直接提供合并后的文件链接也可以，万分感谢。

另：IAPR TC-12 这个数据集可以提供下载地址嘛？

Is that normal？

Hi @jiangqy ,

Thanks for the code. I have a small request, is it possible to share the image net mat file you used in this work?
I think the mat file from here http://www.vlfeat.org/matconvnet/pretrained/ looks different form what you have used.

Thank you.