search

megvii-research / FQ-ViT

[IJCAI 2022] FQ-ViT: Post-Training Quantization for Fully Quantized Vision Transformer
Apache License 2.0
301 stars 48 forks source link

SwinTransformer3D #4

Closed wangjingg closed 2 years ago

wangjingg commented 2 years ago

Does this framework support SwinTransformer3D quantification?

linyang-zhh commented 2 years ago

@wangjingg The motivations of our work is based on the serious inter-channel variation in some layers of transformer architectures, and we propose a series of methods to solve that. Our experiments are tested on 2D transformers such as ViT and Swin. So far, we don't know whether SwinTransformer3D has such characteristic too, so maybe, additional evaluation is required.