penghu-cs/mrl

Learning Cross-Modal Retrieval with Noisy Labels (CVPR 2021, PyTorch Code)

mrl's Issues

Question about symmetric noise rates.

I recently read your paper and very interested in it. When I reading, I'm confused about how to implement "symmetric noise rates". Code of the paper just provides the path of symmetric noise rates JSON file but not provides the detail of it.
So could you share the code detail of add noise? Your help will be greatly appreciated.

top@k precision-recall curve

Hi, I have some questions in your other paper, How to plot a top@k precision-recall curve(e.g,.apply CCA or DCCA method)？ thanks

Hi, it seems that the NUS-WIDE and INRIA-Websearch cannot be downloaded from the given website.

Hi, it seems that the format of these datasets is not consistent with the original data, so could you please send me a copy of the processed data (including the nus_wide_deep_doc2vec_data_42941.h5py and xmedianet_deep_doc2vec_data.h5py)? Thanks!

The retrieval performance measures the instance level or category level?

Hello, thanks for your great work. I'm new at supervised cross-modal retrieval. Given a query sample, what is the correct retrieval?
The retrieval result has the same category as the query sample or the retrieval result must belong to the same instance as the query sample.

Question about Contrastive Learning loss.

To my test, in the function statement as "def cross_modal_contrastive_ctriterion(fea, tau=1.):", values of loss1 and loss2 is qual.
Then, I have been made some qualitative analysis, and find that loss1 and loss2 are equivalent, just the two multipliers of the numerator have swapped positions.
Did the approach is redundant?

About datasets of your experiments.

Dear Peng Hu,

Sorry to bother you again.
There's something I'd like to ask about datasets of your experiments.

In your paper, in addition to "Wikipedia", there were experiments on "INRIA-Websearch", "NUS-WIDE", and "XmediaNet"." The data from "wikipedia" is available on your github, but the data from other datasets are not available. Is it possible to make those datasets available? (i.e., nus_wide_deep_doc2vec_data_42941.h5py, INRIA-Websearch.mat in noisydataset.py, etc.)
If this is difficult, could you please tell me how to create them?

I am sorry for taking up so much of your time but I appreciate your help.

Thanks so much.

Hi, a question about clustering assignments C?

Thanks for your wonderful job and great paper writing.

But I'm confused about why does the performance promote? Because after I read the codes, I realize that the clustering assignment C is just a linear layer like another classification model. It's no different from the normal way. On the other hand, RC loss may be more suitable for noise settings than CE loss, However, its working role is also the same as the CE loss. I guess it's not the key to such a great performance. So, my question is, how does the model deal with noise so efficiently.

Hope for your reply, thanks again.

I have two questions for your paper.

Dear Peng Hu,

I read your paper and it was so impressed. My deepest thanks for providing the code!
I have two questions about your paper.

First, in the Wikipedia dataset, it says that the maximum number of epochs is 100 and the batch-size is 50. However, the github usage says that the maximum number of epochs is 30, and the batch-size is 100. Is it correct to set the maximum number of epochs to 100 and the batch-size to 50?

Second, do the results in your paper report the maximum MAP value throughout all epochs? Or is it the value of MAP in the last epoch?

Sorry to ask this of you when you are busy but I appreciate your help.

Thanks so much.

Hello，can you show the code about how we can get the h5py file from zhe dataset

Question about Robust Clustering loss.

Sorry to bother you again. I find equation (5) in the paper is not similar to the lines 217 in /src/utils.py. Could you explain it. Thanks.