Please note that NO external data (including ImageNet) can be used in both the pretraining and training phase.
All the models should train from scratch since the external information about contexts or categories can help the model learn about the test data which should be totally unseen in the training phase.
Thank you for your understanding!
Please note that when loading Pytorch vision models by default, the ImageNet pretrained weights are loaded as the initialization, which is not allowed in NICO challenge.
So please create model like "self.network = torchvision.models.resnet18(pretrained=False)" as shown in the tutorial.
Thank you for your understanding!
Hey, I am a participant of this challenge.
I just realized that I used 2 constants from ImageNet when training.
They are the default mean and std var of ImageNet.
I use these 2 constants to Normalize my input.
Will this be considered as a violation of the rule that no external data shall be used?
The code I import these 2 constants is:
"""
from timm.data.constants import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
"""
I think we can use the mean and std of this dataset, it may perform better.
Posted by: Iloveconvnext @ June 10, 2022, 3:35 p.m.Hi, the Normalization layer initialization with mean and std from ImageNet is not forbidden since they do not leak the knowledge of target contexts. Yet you could calculate the mean and std of NICO++ training data as the initialization as well.
Posted by: NICO-official @ June 20, 2022, 4:27 a.m.