Reproduce resnet 18 + relabel

[John1231983]

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

[hellbell]

@John1231983
Thanks for having interests in our paper.
Indeed, our results 72.5% accuracy is obtained without CutMix regularization.
We got a slightly better result (72.8%) with the CutMix trick.
It seems that there is a difference in the number of GPUs, we used 4 but 2 for yours, which makes the minibatch size twice than ours.

[John1231983]

We used 2 gpus but bsx2 for each gpu. We also evaluate on x2 batch size and 8 gpu but best accuracy is 72.1. Any suggestion?

[hellbell]

@John1231983
We have run it twice on 4 GPUs and the results were
* Prec@1 72.546 Prec@5 90.478
* Prec@1 72.614 Prec@5 90.432.
Could you run it on 4 GPU setting?

[schallko]

Could you maybe provide the config file of this experiment? I´m struggling to reproduce results aswell.

[hellbell]

@schallko we used the config file for resnet50 training.
Change the architecture to resnet18 at https://github.com/naver-ai/relabel_imagenet/blob/main/configs/relabel_train_resnet50.yaml#L2