Request for Hyperparameters and Synthetic Images for Aircraft and Cars Datasets

[xyzz-jpg]

Hello, thank you for providing and open-sourcing this highly contributive work.

Could you please provide the hyperparameters for the Aircraft and Cars datasets, as well as the synthetic images generated by SD 2.1 for these datasets? I am unable to replicate the results on these datasets.

[goldkim92]

Thank you for your interest.

We are currently working on specifics of if and how to release the dataset.
Meanwhile, you can try with these hyperparameters on both datasets:

DataDream:
    LoRA rank: 16
    learning rate: 1e-4
    epoch: 200

Generation:
    timestep: 50
    guidance scale: 2.0
    
Classifier:
    learning rate: 1e-4
    weight decay: 1e-4
    Mixup & Cutmix: True

[xyzz-jpg]

@goldkim92 Thanks for your reply. However, I still can not reproduce these results.

The hyper-parameters of Classifier I used:

learning rate:1e-4
weight decay: 1e-4
Mixup & Cutmix: True
Epoch: 50
Numbers of synthetic images: 500

The result on aircraft is only 65.16. Hope your help.

In addition, could you please provide training logs on these datasets, especially on aircraft?

[goldkim92]

Could you also provide the DataDream config setting? Also could you tell us whether you used class-wise or dataset-wise setting?

[xyzz-jpg]

Thanks for your reply!

I conducted the experiments under the dataset-wise setting and found that the quality of the generated data seems to be low. Could you please provide me with the hyperparameters used for fine-tuning SD 2.1?