Really a nice work ! But I have some minor questions:

1: I think Eq.5 should be writen as : log E(p,q) + log N = log<Q, -log P>. Can you give some points about how Eq.5 holds on? (althoug it does not influence any conclusion in this paper, I just want to confirm it)

2: After reading your answer about "How do you transition from the probability matrix Q to the labels?" on openreview website, I have some confusion： 1) Do you mean that in step1(representation learning), the Q of eq.6 is actually a one-hot matrix by applying argmax on probability matrix Q^* (the Q^* here means the direct solution of eq.7 in step2)? 2) If I understand correctly, how about use Q^* directly in step1 to compute the cross-entropy loss, but not use argmax(Q^*) ? If a soft label will be better?

p.s. I have asked above questions on openreview website but maybe you did not receive it, so sorry to bother you again here.

Hi

Trying to run the code, I came across the following issues which might be helpful if you fix them:

• Resnet architecture is incorrectly written as "presnet" in scripts/resnet.sh
• In return_model_loader function at data.py, the trainloader is only available for resnet architecture.
• Using only one GPU will result in division by zero multigpu.py
• Using more than one classification head leads to IndexError in CrossEntropy. Could you please check these issues?

Thank you

Thanks for your codebase!

I tried cifar.py for cifar 10 and did not know bellow code.

acc = kNN(model, trainloader, testloader, K=10, sigma=0.1, dim=knn_dim)

kNN devide test data of cifar10 to 10 clusters and calclate acc with GT? If so, how to correspond GT classes and and clusters for eval?

Hi,

You say in a paper 3.5 Since our main objective is to use clustering to learn a good data representation Φ, we consider a multi-task setting in which the same representation is shared among several different clustering tasks, which can potentially capture different and complementary clustering axis

What if my main objective is to receive meaningful clusters. I would like to train model on my own unlabeled, rather messy dataset and then clean dataset from particular images filtering out some clusters. Would you recommend to set number of heads T =1 in that case?

Hi,

I am trying to use your algorithm using another dataset. After few epochs, when I tried to update the pseudo labels, I got this message :

Traceback (most recent call last):
  File "PseudoLabelling.py", line 391, in <module>
    L, cfg.PS = update_pseudoLabels(model, cfg, inferloader, L)
  File "PseudoLabelling.py", line 314, in update_pseudoLabels
    L, PS = cpu_sk(model, pseudo_loader, cfg.hc, cfg.outs, L, cfg.presize, cfg.K, cfg.dtype, cfg.device, cfg.lamb)
  File "PseudoLabelling.py", line 251, in cpu_sk
    L, PS = optimize_L_sk(L, PS, outs, lamb, "cpu", dtype, nh=0)
  File "PseudoLabelling.py", line 306, in optimize_L_sk
    argmaxes = np.nanargmax(PS, 0) # size N
/.conda/envs/torch/lib/python3.8/site-packages/numpy/lib/nanfunctions.py", line 551, in nanargmax
aise ValueError("All-NaN slice encountered")
ValueError: All-NaN slice encounteredes = np.nanargmax(PS, 0) # size N

Why did I get this error ? Is it because every pseudo labels are the same than the previous epochs ?

Hi @yukimasano and @mayank010698 ,

Could you please provide a requirements.txt file for everyone interested to set up the project painlessly? Your README.md suggests the following:

Requirements Python >3.6 PyTorch > 1.0 CUDA Numpy, SciPy (optional:) TensorboardX

Here's a configuration which successfully runs the scripts/eval-resnet.sh script with the resnet-10x3k model:

scipy==1.5.2
tensorboardx==2.1
torch==1.6.0
torchvision~=0.7.0
numpy~=1.19.2
Pillow~=7.2.0

Providing a requirements.txt file can save a lot of debugging hours finding which exact version of torch (> 1.0.0) and torchvision works.

Thank you in advance!

As I know, DeepCluster need re-set classification head https://github.com/facebookresearch/deepcluster/blob/9796a71abbfd14181a2b117d6244e60c2d94efbf/main.py#L145 because there is no mapping between two consecutive assignments.

So in SeLa, if classification head is need to be re-set like deepcluster (It seems I can not find such code)?

If re-set is not need, could you please explain why?

Thanks!

Hello,

I am trying to reproduce the reported results for ImageNet using AlexNet. I have simply cloned the repository and run alexnet.sh script. I am using 4 Nvidia RTX 2080ti GPUs.

In the paper, for the label optimization step, it is stated that "convergence is reached within 2 minutes on ImageNet when computed on a GPU". In my case, it takes 5 days and the program throws an out-of-memory error after that. What kind of GPUs have you used during the training?

Also, as a follow-up: I was wondering if it is possible to share the CIFAR code since it is easier to run and reproduce the results?

Thank you.

In the code snippet, it's written pseudolabels = ckpt['L'] but there's no such key L in the state dict. labels is the correct key I guess.

@yukimasano

Hi, thanks for your great paper. I've noticed that you set shuffle=True in your dataloader t_loader and let train_loader and pseudo_loader point to the same object t_loader. I'm wondering in this way whether the optimised labels are in the right order when they are used to train the model since you will shuffle the data again.

Hi, I wonder if you have the code for getting the pseudolabels of training stage? As you recommend:

"As we show in the paper, the pseudolabels we generate from our training can be used to quickly train a neural network with regular cross-entropy. Moreover they seem to correctly group together similar images. "

I want to get the pseudolabels as I train the model on a new dataset. Thank you!

Hi, I run the cifar.sh in the end, the results output like this:

Epoch: [399][390/391]Time: 0.074 (0.308) Data: 0.008 (0.210) Loss: 1.6734 (1.5705)
10-NN,s=0.1: TOP1:  75.43
best accuracy: 75.62
doing PCA with 128 components ..done
10-NN,s=0.1: TOP1:  75.48
......
Epoch: [399][390/391]Time: 0.074 (0.308) Data: 0.008 (0.210) Loss: 1.6734 (1.5705)
10-NN,s=0.1: TOP1:  75.43
best accuracy: 75.62
doing PCA with 128 components ..done
10-NN,s=0.1: TOP1:  75.48

this result is much less than 83.4 reported in the paper. Can someone tell me why? Thank you very much!

Hello, thanks for your great work. I want to reproduce your results in ImageNet. However, I find that the default learning-rate in code is 0.08, but it is 0.05 in paper. So which learning-rate I should use to reproduce the results?