Code for Imagenet about agem HOT 6 CLOSED

We do not provide the script for mini Imagenet. Once you download the mini Imagenet dataset, you can edit the conv_split_cifar.py file to run the mini Imagenet experiments. The setup is the same as that of Split CIFAR; a total of 100 classes split into 20 tasks with 5 classes per task. The hyper-parameters for mini Imagenet are given in the appendix of https://arxiv.org/abs/1902.10486.

from agem.

Thanks for your answer. I would like to make sure my preprocessing is the same as yours. I see that you have shared code to load imagenet, all that's missing is the .pkl file you used. Could you point me towards the download link you used (or upload the .pkl file to the repository) ?

from agem.

You can download the miniImageNet dataset (pickle file) from here: https://www.dropbox.com/s/yt3akdfchuafk25/miniImageNet_full.pickle?dl=0

Hope that helps!

from agem.

Thanks for sharing. It seems there was some preprocessing done from the link you shared (1 pickle with 84 x 84 images) to the one used in your code here (4 pickles, 224 x 224 images, train and test already split).

I guess the only information I'm missing now is

1. how to split train and test
2. if you could confirm you did resize to 224 x 224 (and if you used data augmentation)

from agem.

The code for imageNet utils in the repository is outdated. I used 84 x 84 images. The train/ test split is 500/ 100 for each class. You can modify the cifar utils for imageNet.

from agem.

great! closing the issue

from agem.