Please provide the location/weight of ViT-G/14 trained model on ImageNet?

[s-kumar17]

No description provided.

[andsteing]

We never trained a ViT-G/14 on ImageNet.

Note that already in the original paper https://arxiv.org/abs/2010.11929 we didn't train anything larger than ViT-L/16 on ImageNet (see Table 5). That is because large models require more data other things being equal (see Figure 3 from that same paper).

Things look a bit better when you add data augmentation / model regularization. See the "AugReg" checkpoints from https://arxiv.org/abs/2106.10270 that we published in this repository. But here we were using ImageNet-21k, which is a lot larger than ImageNet, and didn't go beyond the L/16 model (this approach might work for a larger H/14 model, but we haven't trained that).