search

TACJu / TransFG

This is the official PyTorch implementation of the paper "TransFG: A Transformer Architecture for Fine-grained Recognition" (Ju He, Jie-Neng Chen, Shuai Liu, Adam Kortylewski, Cheng Yang, Yutong Bai, Changhu Wang, Alan Yuille).
MIT License
382 stars 88 forks source link

train from scratch #16

Open MengXinChengXuYuan opened 3 years ago

MengXinChengXuYuan commented 3 years ago

Did anyone try to train from scratch, without any pertained weight? I want to make the model adapt to my project, with 224 224 input, 8 8 patches and 6 sliding size, which means there is no pertained weight for me. I found it very hard to converge, after 10000 steps the train acc is still around 0.6 Than I tried the original training configurations, except loading pertained weight, the same issue Did I miss anything?

qing-xue commented 2 years ago

Hey, I got the same problem, and my best acc is just 0.8.

bjtuln commented 2 years ago

I got the same problem