Which type of pre-trained ResNET was used?

[jihwankwak]

Hello thanks for your hard work :)
I read the ICML 2019 version of your paper and I came up with some questions about implementation of pre-trained ResNET-18 model.

FIrst ot all, is pre-tranining for ResNET-18 on ImageNET omitted in your code?
I saw in the paper that the pre-training was done on ImageNET which only has non overlapping classes with CIFAR100. I was looking for your pre-training implementations (Augmentation, epochs, lr, and etc.) in the paper and code (Maybe it might be my bad) but, currently, I am having hard time finding it.
Also, below line in your code using pretranied model from torchvision even confuses me.

lifelong-learning-pretraining-and-sam/img_exps/vision_utils.py

Lines 54 to 62 in 2fee18a

	def __init__(self, total_classes, classes_per_task, pretrained=False):
	super(ResNet, self).__init__()
	resnet = torchvision.models.resnet18(pretrained=pretrained)
	for param in resnet.parameters():
	param.requires_grad = True
	self.net = resnet
	self.net.fc = torch.nn.Linear(resnet.fc.in_features, total_classes)
	self.classes_per_task = classes_per_task

So, can you specify your pre-training implementation of your ResNET-18 model on ImageNET with less classes?

Thanks again :)