yitu-opensource / T2T-ViT

ICCV2021, Tokens-to-Token ViT: Training Vision Transformers from Scratch on ImageNet

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Questions about the results of the experiment

chenyucong1 opened this issue · comments

Hi author, why is the accuracy and loss of test so different from the data of EMA?
微信图片_20210330153816

Hi, it seems that the '--model-ema-decay' is not suitable.
So you can:
(1). disable --model-ema;
(2). change the value of '--model-ema-decay' based on your iteration number in training.