how to use this model to train and evaluate on Imagenet?

[Alexanzhuo]

trafficstars

[Alexanzhuo]

Thank you for providing this network! I want to use this model "ViT" to classify Imagnet,but the accuracy is not good.I try it on mini-Imagenet first. I use the network like this: net = ViT( image_size = 224, patch_size = 16, num_classes = 64, dim = 1024, depth = 6, heads = 8, mlp_dim = 2048, dropout = 0.1, emb_dropout = 0.1 ).cuda(device)

[Alexanzhuo]

But the accuracy just stops increasing.

[lucidrains]

@Alexanzhuo Hi Alex, you won't see any positive results in the small data regime. What you can do, however, is to run self-supervised learning (BYOL) on a bunch of unlabelled images first, and then train on your tiny mini-Imagenet corpus.

Or you can just use Ross' version with the pretrained weights released by Google