search

Liangqiong / ViT-FL-main

MIT License
103 stars 18 forks source link

Pretrained models on ImageNet21k vs ImageNet1k #1

Closed cbbjames closed 3 years ago

cbbjames commented 3 years ago

Hi,

Thanks for the interesting work!

May I confirm that are the results of ResNet50 in the paper are pretrained on ImageNet21k as the same as ViT? Or is there a ResNet50 ImageNet21k pretrained model that needs to be used here (instead of the ImageNet1k one from official PyTorch)? https://github.com/Liangqiong/ViT-FL-main/blob/7521a2ae915747d53cc1de287e852a10002895d4/utils/start_config.py#L73

Thanks!

Liangqiong commented 3 years ago

The current version here we take the default pretrained models, for ResNet use Image1k, and ViT use Imagenet21k. We have a new version that also evaluate the VIT with pretrained model on Image1K, but not published yet.

cbbjames commented 3 years ago

Thanks!

Liangqiong commented 2 years ago

Just an update, the current version, all the pretrained models are from ImageNet1k