Hi, thanks for sharing the great work! I found that the same Evaluation Metric, i.e., the Hungarian algorithm, is used for seen classes and novel classes. In fact, I think that the standard "classification accuracy" should be used for seen classes, to avoid the mis-matching issue on seen classes.

In multiple places in the codebase, the mask to identify one class comes from the old or new classes is computed as:

• https://github.com/sgvaze/generalized-category-discovery/blob/262519540c3452c6b8276dc203b4e9067378c552/methods/clustering/k_means.py#L48-L49
• https://github.com/sgvaze/generalized-category-discovery/blob/262519540c3452c6b8276dc203b4e9067378c552/methods/contrastive_training/contrastive_training.py#L313-L314
• https://github.com/sgvaze/generalized-category-discovery/blob/262519540c3452c6b8276dc203b4e9067378c552/methods/estimate_k/estimate_k.py#L49-L50
• https://github.com/sgvaze/generalized-category-discovery/blob/262519540c3452c6b8276dc203b4e9067378c552/methods/estimate_k/estimate_k.py#L116-L117

However, this is only compatible with the setting that all old classes are consecutive (e.g., 0, 1, ..., n_old_classes - 1). For the SSB splits where the old class indexes are non-consecutive, it can result in incorrect old-new splits.

Influence of this bug: the old/new ACCs of CUB & Stanford Cars are not reliable (if use_ssb_splits==True), but the all ACC is not affected.

Possible fix: replace range(len(args.train_classes)) with args.train_classes.

Dear Sagar Vaze, I am reading your released code of your interesting paper Genearalized Cateogory Discovery. However, I seem to be confused about one implementation in the file 'contrastive_training.py'. In your function 'test_kmeans.py', when doing the KMeans, why the @n_clusters equals to the ground-truth class number here? Do you assume here you have access to the ground-truth cluster number or I mistakenly understand the whole pipeline? Kindly could you help to explain this?

Thanks for any help you can give.

Hi, thanks for sharing.

But I met a problem that is I can't the file: dino_vitbase16_pretrain.pth, from the project folder. Could you please upload this file?

Thanks a lot.

Hi, Thank you for providing the code for your interesting work. I am trying to reproduce the results on imagenet-100, however, the class ids are not provided in this repository. In data/imagenet.py it seems class selection is random. Can you please share the 100 class ids(Imagenet folder names ) that you used to generate your results?

Hi, I was looking into the dataset for Herbarium19. After downloading the dataset , I can only see 3 splits :'train', 'validation' and 'test'. May I know what are the 'small_train' and 'small_validation' which you used in herbarium_19.py ? Also, each training epoch for herbarium19 takes a very long time to run. Is it the reason that you used a subset of it?

Hi, great work.

May I know if table 2 and 3 are based on clustering method mentioned in 3.1.2 or 3.2?

The numbers in table 3, specfically on Herbarium19 shows (35.4 | 51.0 | 27.0) which is the same as table 5 (6), which suggest semi-sup k-means was used to generate the results for table 3. While, table 2 results on CIFAR100 shows (70.8 | 77.6 | 57.0 ) which is different from table 5. Thanks in advanced.

Hi, Thanks for releasing the code for your interesting work. I am trying to run semi-supervised kmeans with bash bash_scripts/k_means.sh . but it seems to be pretty slow. Is this expected ?

Patching CVE-2007-4559

Hi, we are security researchers from the Advanced Research Center at Trellix. We have began a campaign to patch a widespread bug named CVE-2007-4559. CVE-2007-4559 is a 15 year old bug in the Python tarfile package. By using extract() or extractall() on a tarfile object without sanitizing input, a maliciously crafted .tar file could perform a directory path traversal attack. We found at least one unsantized extractall() in your codebase and are providing a patch for you via pull request. The patch essentially checks to see if all tarfile members will be extracted safely and throws an exception otherwise. We encourage you to use this patch or your own solution to secure against CVE-2007-4559. Further technical information about the vulnerability can be found in this blog.

If you have further questions you may contact us through this projects lead researcher Kasimir Schulz.

Thanks for your brilliant work! I wonder that whether the binary search function is valid. As known to me, the binary search can be utilized to estimate the minimum of function by searching the root of f'(x)=0, but only f(x)=ACC is given while f'(x) is not given. So I have some questions on it and hope to receive your response to it.

Thanks for the interesting research. Are the k values ​​used in the benchmarks(tabel 2,3) estimated? or GT? In my opinion, there will be a difference in acc depending on the k value, but in the code, it seems that the performance was measured through GT. What results were used in the paper? Also, is there a result of measuring the change according to the k value?

Reimplementing these two methods in the GCD setting is pretty cumbersome, as there are too many details to align. It would be greatly helpful if you could provide the code for implementing them or the trained models' checkpoints (or the models' predictions) for evaluation. Many thanks!

I ran the experiment of table 2 using this repo and the final accuracy for all classes on CIFAR10 and CIFAR100 are only 60.28 and 38.60, respectively, which are much lower than the reported.

To reproduce this paper, I ran the following three files sequentially: "contrastive_train.sh", "extract_features.sh", and "k_means.sh". Following the implementation details in the original paper, I used ViT-B-16 backbone with DINO pre-trained weights (weights are downloaded from this repo: https://github.com/facebookresearch/dino) and fine-tune the final transformer block. The total number of clusters for k-means clustering is set as the total number of classes.

The default setting of this repo seems to be set on Stanford Cars, so can you give detailed parameters for CIFAR10/CIFAR100??

Are the experiment results in table2 and table3 performed on unlabeled data of the training set？ I found that k_means.py in your code to do just that?