Support validation without create the training dataset and dataloader

BR-IDL / PaddleViT

:robot: PaddleViT: State-of-the-art Visual Transformer and MLP Models for PaddlePaddle 2.0+

https://github.com/BR-IDL/PaddleViT

Apache License 2.0

1.22k stars 318 forks source link

Support validation without create the training dataset and dataloader #115

Closed xperzy closed 2 years ago

xperzy commented 2 years ago

Describe your feature request Current validation (config.EVAL is True) mode still create and load training dataset and dataloader, which is not flexible when users only have val set.

So main method needs support val mode without touching training set.

Describe the reference code or paper N/A

Describe the possible solution I have a fix in ViT model, which can be used to other classification model. Please refer to commit 9a7c105 for details.