naver-ai / relabel_imagenet Goto Github PK

Is there a reasoning script, can you view the results?

Hi,

This is exciting work! Would you be able to please provide the code that you used to generate the relabel imagenet maps? I would like to apply this technique to my own dataset and the relabeling code would be super helpful. Even if it's not polished, it would be a great help :)

Thanks,

Could you release the code which is used to relabel imagenet ?

if i use our data,do not use ImageNet data,so,how do we get Relabel data?

Thank you for the inspiring work! I have read the paper and the code, and would like to raise a question about the selection of the training hyper-parameters. Specifically, I found that you use custom setups for optimizers (e.g., SGD v.s. AdamP, learning rate, weight decay, etc.) in three configs (baseline, relabel, relabel+cutmix). I am wondering that how did you tune it? Are there any policies? Thanks!

Thank you very much for your meaningful research work!

For 4.3 Multi-Label Classification, I would like to ask where is the code for LabelPooling?

A sample of relabel dataset has shape of [2,5,15,15]

Could you explain what each axis means?

I believe that [0, :] means prediction confidence and [1:,] means predicted class number.

https://www.dropbox.com/s/9sxigpec7fxq8wh/relabel_imagenet.tar?dl=0 여기 링크에서 데이터 다운로드 계속 실패인데..혹시 다른 방법 더 있을가여??

paragraph

def get_relabel(label_maps, batch_coords, num_batches):
    target_relabel = roi_align(
        input=label_maps,
        boxes=torch.cat(
            [torch.arange(num_batches).view(num_batches,
                                            1).float().cuda(),
             batch_coords.float() * label_maps.size(3) - 0.5], 1),
        output_size=(1, 1))
    target_relabel = torch.nn.functional.softmax(target_relabel.squeeze(), 1)
    return target_relabel

line
batch_coords.float() * label_maps.size(3) - 0.5
batch_coords is given by percentage [x0,y0,x1,y1],
and it recovers to region [0,15].
why coords is substracted by 0.5?

Could you provide another source of relabeled-dataset? Dropbox really annoys me when it breaks.

Hi, I reproduced your paper with resnet 18 and 2A100 GPU, batch_size: 1024 + cutmix, but my result is not good (about 72.004%)

Epoch: [299][250/626]   Time 1.088 (1.088)      Speed 1882.005 (1882.005)       LR 2.05E-08     Loss 2.3402 (2.3402)    Prec@1 66.016 (66.016)  Prec@5 86.719 (86.719)  
Epoch: [299][500/626]   Time 1.103 (1.096)      Speed 1856.418 (1869.124)       LR 2.30E-09     Loss 2.2720 (2.3061)    Prec@1 69.775 (67.896)  Prec@5 88.477 (87.598)  
[Epoch 299] 780.568 sec/epoch   remaining time: 0.000 hours
 * Prec@1 72.004 Prec@5 90.296

Your report is 72.5 %

ResNet-18 | 71.7 | 72.5 (+0.8) [model_file]

Any suggestion to reproduce your result? Thanks

Hi
nice work, i enjoyed reading the article.

i noticed that you are using "plain" cutmix in the training.
have you considered doing "relabeld-cutmix", meaning pooling the targets from the relevant area only ?

it is definitely more complicated, but I think it can even count as a new interesting type of augmentation, that can be used
not only when learning from a teacher

Tal

Oh no. I thought i could use the relabeled dataset to test my model, so i download imagenet val in 10GB.
But it is that the relabeled is for the training split which is 100GB.

Do you have relabeled val? I am doing the downloading.

BTW. Is ImageNet 2012? Does train and val overlap in samples?