sail-sg / poolformer

PoolFormer: MetaFormer Is Actually What You Need for Vision (CVPR 2022 Oral)
https://arxiv.org/abs/2111.11418
Apache License 2.0
1.3k stars 117 forks source link

s12 model Reproduction experiment #42

Closed starsky68 closed 1 year ago

starsky68 commented 1 year ago

Using the s12 model, only the four card batch size is 240 for a single card, and the acc top1 is 76 in the end. If there are no eight cards, how can the acc reach 80,Other parameter defaults. --Apex amp can greatly affect the accuracy in addition to fast training.

yuweihao commented 1 year ago

Hi @starsky68 ,

Thanks for your attention. Please refer to the train.py file in metaformer repo where I add --grad-accum-steps so that you can set larger batch size.