Can't Reproduce Result in CIFAR-20

[OBJECT907]

Can't Reproduce Result in CIFAR-20

[Hzzone]

You may need to modify config/cifar20_r18_byol.yml with the changing of parameters cluster_loss_weight and latent_std. I am running this setting, I will upload the pre-trained model and give you feedback as soon as possible.

[OBJECT907]

You may need to modify config/cifar20_r18_byol.yml with the changing of parameters cluster_loss_weight and latent_std. I am running this setting, I will upload the pre-trained model and give you feedback as soon as possible.

Thanks.

[Hzzone]

This issue has been solved, and I have updated the pre-trained model in the download link and achieved the same results in my paper.

[OBJECT907]

What training parameters need to be changed if I use resnet34?

[Hzzone]

Hi, I do not understand what you mean. Could you detail the problem?

[OBJECT907]

I simply change encoder_name to bigresnet34 in config/cifar10_r18_propose.yml but I got acc 53% (even worse than resnet18).
Is there any other config I need to change?

[Hzzone]

Have you changed the num_devices? I will check if the time is available.

[OBJECT907]

I trained Propos on 4 RTX 3090 GPUs. num_devices is 4.

[Hzzone]

You may need to change config/cifar20_r18_propos.yml. The training looks like this:

[OBJECT907]

You may need to change config/cifar20_r18_propos.yml. The training looks like this:

This is not the result of CIFAR-20, I guess you upload wrong image. QwQ

[Hzzone]

No, it is the result of R34 on CIFAR-10.

[OBJECT907]

So how can I train Propos on CIFAR-20 with resnet34?

[Hzzone]

I do not understand which datasets you would like to train, since you mentioned you have changed the config/cifar10_r18_propos.yml; that is why I post the training on CIFAR-10. In addition, as I have said above, if you need to train on CIFAR-20, plz change config/cifar20_r18_propos.yml.

[OBJECT907]

Sorry. It's my typo. I want to reproduce your result on CIFAR-20 with resnet-34. Do I just need to change the encoder_name in config/cifar20_r18_propose.yml ?