Closed NilakshanKunananthaseelan closed 2 months ago
Thank you for your interest in our work. We will endeavor to update the code as soon as possible.
It seems that the model you trained has not converged and thus cannot be used as an origin model for unlearning. We recommend using a larger learning rate and training for more epochs. For instance:
python main_train.py --dataset TinyImagenet --batch_size 256 --epochs 160 --lr 1e-1
If you have any other questions, please feel free to contact us.
Issue closed. Please feel free to post new responses or initiate another issue if you have further questions. Thanks!
Thanks, @a-F1. I'll use this command and update you.
It seems that the previous command is causing overfitting. Have you used any additional regularization other than the default settings?
Hi @NilakshanKunananthaseelan, the normal trained ResNet-18 on TinyImagenet is ~60 testing ACC. If you want to pick the highest ACC, feel free to stop early.
Hi Team, Appreciate your work. I'm interested in unlearning methods for higher-resolution images. CIFAR10 models are ~100% accurate, this makes me curious if are we unlearning correctly. I couldn't find
"Maysee/tiny-imagenet"
files, therefore I downloaded"zh-plus/tiny-imagenet"
. I don't see theimagenet.py
creating the directory structure for loading the dataset withImageFolder
. Therefore, I used the following snippet to create it.And used the following command to train the ResNet18 model.
python main_train.py --arch resnet18 --dataset TinyImagenet --lr 0.01 --epochs 80 --save_dir ${save_dir} --data_dir ${data_dir}
The learning saturates quickly in 30 epochs with the best accuracy of around 56%.![image](https://github.com/OPTML-Group/Unlearn-Saliency/assets/25349618/88e077d1-8d06-466d-874a-599e54f1bd86)
Can you also release the code related to supplementary experiments?
Thanks