search

NVlabs / SegFormer

Official PyTorch implementation of SegFormer
https://arxiv.org/abs/2105.15203
Other
2.55k stars 353 forks source link

the pre-trained weight of Mit b0 - b5 #151

Open xiuguangLi opened 3 months ago

xiuguangLi commented 3 months ago

Hello, it seems that the pre-trained weights of mit b0-b5 are required when segformer is trained in the code, could you please share the download link for them? Thanks! I did not use the mmsegmentation framework. I extracted the network structure of segformer from your code and verified it with the weight file you provided. There was no problem. However, when using this network to train segformer from scratch on the ADE20k dataset, the results are very low, and it is not known if the reason is that mit backbone's pre-training weights are not loaded.

ywxhjn commented 1 week ago

have you got it, i need it also

xiuguangLi commented 1 week ago

mit权重 https://www.alipan.com/s/VxcinMdznzA 提取码: 38fw 点击链接保存，或者复制本段内容，打开「阿里云盘」APP ，无需下载极速在线查看，视频原画倍速播放。 这是我找到的 mit 的权重，你可以验证一下。 ---- Replied Message ---- @.>Date10/11/2024 @.>@.>, @.>SubjectRe: [NVlabs/SegFormer] the pre-trained weight of Mit b0 - b5 (Issue #151) have you got it, i need it also — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>