Closed knaffe closed 3 years ago
DeiT looks cool! However, this project is focused on providing a Keras implementation of the ViT models from the official repository and so a DeiT implementation would be out-of-scope. The ViT weights in this repository are the same files provided by the official repository. They're copied into the repository only for convenience and long-term reliability (i.e., in case the original files get deleted or moved). The vit_keras
package just loads those NPZ weights into a Keras model using vit_keras.utils.load_weights_numpy
, which works only with the ViT weights. It's not going to work with weights from any other source.
Thanks for your great contribution! I notice you offer the VIT npz format models. How could you convert official models from JAX to NPZ?
I want to convert the Deit official models and apply into your repo. Thanks a lot again.