Hi, thanks for your great work! I notice you use the cosine classifier in many experiments and it can get a better baseline. The formula is as follows

I am wondering the value of s?

Hello. Thank you for your project! I'm testing your code on my custom dataset. My task is classification. I have a question about your code implementation.

https://github.com/Megvii-BaseDetection/DisAlign/blob/a2fc3500a108cb83e3942293a5675c97ab3a2c6e/classification/imagenetlt/resnext50/resx50.scratch.imagenet_lt.224size.90e.disalign.10e/net.py#L56-L62

From my understanding, in stage 2, remove the linear layer used in stage 1 and add DisAlign Linear Layer. And freeze all parts except for logit_scale, logit_bias, and confidence_layer. At this time, the weight and bias of DisAlignLinear are also frozen. (self.weight, self.bias) Is my understanding correct?

If so, are the weight and bias of DisAlignLinearLayer fixed after the initialization? (The weight and bias of the linear layer in stage 1 are not copied either)

If my understanding is correct, why is the weight of DisAlignLinear also frozen?

I will wait for your reply. thanks!

Hello. Thank you for your impressive project!

I want to apply DisAlign to classification. However, an error occurs in the import part. https://github.com/Megvii-BaseDetection/DisAlign/blob/a2fc3500a108cb83e3942293a5675c97ab3a2c6e/classification/imagenetlt/resnext50/resx50.scratch.imagenet_lt.224size.90e.disalign.10e/net.py#L7 I coudn't find the DisAlignLinear in cvpods.layers. and there also isn't exist at https://github.com/Megvii-BaseDetection/cvpods/tree/master/cvpods/layers How can I solve this problem?

Thank you!

I tried to train the proposed method on ImageNet_LT, but I can only get an average testing rate about 49%, which is far from the rate described in the paper (52.9). Some of the details regarding my implementations are given as follows: (1) The feature extractor is ResNexT-50 and the head classifier is a linear classifier. The testing accuracy in Stage-One is 43.9%, which is OK.

(2) The testing accuracy of adopting cRT method in Stage-Two is 49.6%, which is identical to one reported in other papers. (3) When fine-tuning the model in Stage-2, both the feature-extractor and head classifier are frozen, and a DisAliLinear model (which is implemented in CVPODs) is retrained. The testing accuracy can only reach 48.8%, which is far away from the one reported in your paper.

Hi, thank you for the nice work, but the code for semantic segmentation is missing and the URL for it in the README could not be opened. Could you please fix this issue?

Hi, Thank you for providing your code. Here I was wondering the Equation (10) in your paper (The definition of p_r), which seems not to be a distribution. Since every x_i can only have one label, the reference distribution p_r(y| x_i) will be the distribution like (0, 0, 0,...,w_c, 0, 0,...,0). And the sum of this distribution is w_c, but not 1.

Could you help me understand this equation? Thanks in advance.

Hi, thanks for the great work. Maybe I missed it, but it seems that the code for this project has been incorporated into cvpods. I couldn't launch any experiments due to ImportErrors like: from cvpods.layers import DisAlignLinear ImportError: cannot import name 'DisAlignLinear' from 'cvpods.layers' Also, I didn't find the corresponding functions in cvpods.

Any help will be appreciated. Thanks.

In the paper, the σ(x) is implemented as a linear layer followed by a non-linear activation function (e.g., sigmoid function) for all input x. How to understand the input x？the matrix of raw iamge, or the extracted features, even or cls_score? Thank you!

Dear author, I have some questions about the code and paper:

1. exp_reweight = exp_reweight / np.sum(exp_reweight) * num_foreground Why "exp_reweight" is multiplied by the coefficient "num_foreground"? It is not mentioned in the paper.
2. Is "K" in the empirical class frequencies r = [r1, · · · , rK] on the training set in the paper the same as the class number C of the training set?

Hi, thank you for your amazing work. I want to have a try on semantic segmentation problem. But the link https://github.com/Megvii-BaseDetection/DisAlign/blob/main/TODO of DisAlign_Seg cannot open. Could you pls have a look? Thank you.

https://github.com/Megvii-BaseDetection/DisAlign/blob/main/classification/imagenetlt/resnext50/resx50.scratch.imagenet_lt.224size.90e.disalign.10e/config.py#L31

It seems that you only use test dataset? What is the reason for that?

Thank you for your great work. I met the question of 'AttributeError: 'ConfigDict' object has no attribute 'model' in testing the trained semseg DisAlign model when running the file https://github.com/Megvii-BaseDetection/DisAlign/blob/main/semantic_seg/exps/ade20k_fcn_disalign/disalign_fcn_r50-d8_512x512_160k_ade20k.sh. Is this a version problem of mmcv-full or something else? And can you perfect the explanation of the disalign part of Readme.md in semantic_seg?

Hi, I have a questation about the logit_scale and logit_bias in semantic_seg. The shape of the above parameter is (1, num_classes, 1, 1), why not is (1, num_classes, 512, 512) which is matched the input image size for semantic segmenation.

Hi, did you check the value change of the learned scale and bias vector throughout the training process? I find the value of them change in the first few iterations and remain stable in the rest time on my own classification dataset. I wonder how the learned vectors look like in your paper? Thanks!