hello. very sorry for commenting here, my question is from ppn portfolio which its issues is closed and I had no other way to ask from you. i tried to run that project on google colab but it took so much time which colab doesn't accept. i figured out that tensorflow 1.4.0 doesn't use gpu. is there any solution for that? i tried so much but i got no answer. please help me. and if there is anything that i should consider while using colab for that project, please remind me.

regards

Hi,

Your excellent work really catches my eye, and I want to try TADE on my own dataset to test if it works for industry tasks, but the results doesn't look good compared with conventional methods like focal loss. Results is shown below:

TADE without Test Training: tr: 84 acc val: 87 acc TADE with Test Training 1 epoch: ** val 82.25** TADE with Test Training 5 epoch: ** val 38.41** TADE with Test Training 8 epoch: ** val 30** Focal Loss:tr: 88 acc val: 89 acc

It seems like that result of TADE without test-training is slightly worse than Focal Loss. And with the increase of Test-Training epochs, the accuracy becomes worse. The custom dataset is about a industrial defect classification task, so most pictures have similar background. And pictures can be divided into three categories. train datasetset cls_num_list = [2883,1019,56]

Question: I am not sure if it is because there are only three categories in my dataset, so that it is hard for the output_logit vector to represent similarity, which makes the performance of self-supervised aggregation worse. Do you have idea about that?

Hello， In your paper,the top-1 accuracy on CIFAR10-LT(Imbalance Ratio=10,100) is 90.8% and 83.8%,but when I run your source code,the top-1 accuracy on CIFAR10-LT(Imbalance Ratio=10,100) is 90.16% and 82.92%,What are the specific Implementation details on CIFAR10-LT? Thank you~

Are you offsetting the batch_size for the number n_gpus in the config itself?

CIFAR100:

• n_gpus = 1
• batch_size = 128 Effective batchsize => 128*1 = 128 ?

ImageNet-LT:

• n_gpus = 2
• batch_size = 64 Effective batchsize => 64*2 = 128 ?

iNaturalist18:

• n_gpus = 4
• batch_size = 512 Effective batchsize => 512*4 = 2048 ?

1. For CIFAR100-LT a. Are there different val and test set? b. On what dataset split do you choose the best-trained model? c. What split do you use for hyperparam tuning?

2. For iNaturalist18 a. Are there different val and test set? b. On what dataset split do you choose the best-trained model? c. What split do you use for hyperparam tuning? d. Even though there is an officially available test set (https://github.com/visipedia/inat_comp/tree/master/2018#Data) for iNaturalist18, why don't you use that?

3. General doubts a. What seeds do you use? b. Do you take a mean of multiple seeds?

A great job. Your work solves a wider range of LT problems.

But I m confused with TADE performance on the vanilla LT test set.

Actually, with the same backbone and training strategy, the following methods adopt almost the same loss, but the top-1 ACC varies, for example on CIFAR100-LT-IR-100:

• ICLR'21 logit adjustment [43.89% cf. origin paper Tab.3 ]
• CVPR'21 LADE without test prior [45.6% cf. this paper Tab.8(a)]
• NeurIPS'20 Balanced Softmax which can be rewritten as Eq.3 in this paper [46.1% cf. this paper Tab.8(a)]

In such a situation, TADE should get the best performance when the expert E2 (Eq.3 in this paper) mainly works. If so, it should not outperform the above methods by a large margin, right?

However, the TADE's top-1 ACC is 49.8% (cf. this paper Tab.8(a)) and the weight of experts is [0.40 0.35 0.24] (cf. this paper Tab. 12). The E1 mainly works.

So I just wondering how to explain the improvement of TADE on the vanilla test dataset?

Hi, Thank you very much for your work. I would like to ask if you tried to use ResNeXt101-32x4d instead of ResNeXt50 in your experiments. After my experiments, ResNeXt101 is not as effective as ResNeXt50. Are there any other parameters that need to be changed besides the backbone?

Best,

In line 200 and 201 of test_training_cifar.py: dataset = IMBALANCECIFAR100(data_dir, train=True, download=True, transform=train_trsfm, imb_type=imb_type, imb_factor=test_imb_factor, reverse=reverse) train_dataset = IMBALANCECIFAR100(data_dir, train=True, download=True, transform= TwoCropsTransform(train_trsfm), imb_type=imb_type, imb_factor=test_imb_factor, reverse=reverse) why you set the train is True? I think it should be False to obtain the weighting parameters of test set. Can you explain it? Thanks!

Hi~ Thanks for your excellent work. I have two questions about the paper and the code. (1) I notice that, in default, the shared backbone contains only the layer_1 & layer_2 of resent, other layers in resnet (layer_3 & layer_4) are all in the "expert". Could this setting be described as "shared backbone"? I mean, by saying "shared backbone", the readers will assume that only the classier layer is treated as the "expert". (2) Have you tried the setting that only the classier layer is treated as the "expert"? How much the performance decrease is?

In 4.3 of the paper, objective function is used to calculate the weight of each expert, I guess it is in test_all.py, but I can't find it. Please tell me the spefic position in the code, thank you!

Hi Vanint, I notice that in your LDAM loss implementation the scale is applied on the the adjustment only

x_m = x - batch_m * self.s 

which is different from the original LDAM loss

 return F.cross_entropy(self.s*output, target, weight=self.weight)

basically equivalent to

x_m = (x - batch_m) * self.s 

could you explained more on this detail? Is the coefficient absorbed somewhere on the logit output?