search

kytimmylai / NoisyNN-PyTorch

non-official NoisyNN Implemnentation
Apache License 2.0
48 stars 2 forks source link

About NoisyNN pretrained weights #2

Closed pavaris-pm closed 7 months ago

pavaris-pm commented 7 months ago

Since this is an unofficial implementation of NoisyNN, which means that we do not have a pretrained model weight. All we have is an estimated architecture. If we need to achieve the nearest result according to the paper. We need to pretrained it with the full pretrained dataset right?

Or you already made some experiment from partial samples of it? Is it work ?

kytimmylai commented 7 months ago

Hi,

The pretrain weight is usually inherited from the previous work. We typically don't retrain it.

I care about the 95% accuracy on Imagenet w/ ViT-B, so other experiments may not be considered.