Some question about trainning

[zbw0329]

Hi~Thanks for your excellent work!
I have a machine with 2 1080Ti,and I want to train your model on CIFAR10 with resnet18.

I use the parmeters like this ,but it seems don't work.

The program is stuck in this situation.

[zbw0329]

If I use this parameters

And I get this back

[akuxcw]

Hi, if you want to train models on CIFAR10, you have to change the dataset and data augmentation. The methods of loading ImageNet and CIFAR10 are different.

[zbw0329]

What is the different between '--resume' and '--pretrained'?
How to evaluate my model?

[zbw0329]

I changed the function of loading datasets,it works.

But when I evaluate my model,the output tensor is [1,1000] and the target tensor is [1].
I think it cause by the difference between CIFAR10 and ImageNet.
Do you think I should add a softmax to the end of the model?

[akuxcw]

What is the different between '--resume' and '--pretrained'?
How to evaluate my model?

"resume" will load optimizer, which is used to resume training from unexcepted break.
You can use linear evaluation or downstream tasks to evaluate the model.

[akuxcw]

I changed the function of loading datasets,it works.

But when I evaluate my model,the output tensor is [1,1000] and the target tensor is [1].
I think it cause by the difference between CIFAR10 and ImageNet.
Do you think I should add a softmax to the end of the model?

I suppose the dataloader is wrong. You do need to change the model's classifier to fit CIFAR10, but the shape of output and target tensor is wired. Do you use batch_size = 1?

[zbw0329]

Yes, I set batch_size to 1 by mistake.
Thanks for your help.
I wish you success in your research.

[zbw0329]

My Loss_clu is always the same during my trainning,what happened?

[zbw0329]

Could you release the code of CIFAR10?
I notice that you have showed the table of CIFAR10 on your paper

[akuxcw]

Hi, the results on CIFAR are linear evaluation / finetune results. We didn't try to train pretrained models on CIFAR. The losses during training do reduce very slowly. Are the linear evaluation results on CIFAR reasonable?