rulixiang / ToCo

[CVPR 2023] Token Contrast for Weakly-Supervised Semantic Segmentation
148 stars 11 forks source link

使用vit_base_patch16_224backbone #14

Closed pipizhum closed 1 year ago

pipizhum commented 1 year ago

为什么使用vit_base_patch16_224精度很低?怎么使用vit_base_patch16_224预训练权重?

rulixiang commented 1 year ago

你好,你未提供具体信息,我无法判断原因。 使用vit_base_patch16_224只需将backbone参数改成vit_base_patch16_224。

hly1999 commented 1 year ago

Hellow, I have the same question, I change backbone to vit_base_patch16_224, but the result is always low than report , how to deal with it?

zwyang6 commented 9 months ago

https://github.com/rulixiang/ToCo/blob/e13514d9a89a6c0b4131392a465fcf368a562b4e/model/backbone/vit.py#L237

Hello, Thanks for your contributions! I am wondering why the gradient of the position embedding in the provided vit is off ? shoudln't it be optimized? `

ZhaofengSHI commented 8 months ago

我遇到了一样的问题,发现是因为由于timm版本原因,vit预训练权重没有自动下载导致没有成功载入,把timm版本换成0.5.4就好了