Visual-Attention-Network/SegNeXt

ImageNet pretrainning configurations

rayleizhu opened this issue · 3 comments

rayleizhu commented

Hi, can you provide ImageNet pretraining configs (e.g. resolution, epochs, training tricks)? They are important for a fair comparison.

MenghaoGuo commented

Hi, SegNeXt follows VAN for ImageNet pretrain.

Please refer to this repo.

rayleizhu commented

Hi, SegNeXt follows VAN for ImageNet pretrain.

Please refer to this repo.

This is still vague. For example, VAN has both 224 resolution pretraining and 386 resolution pretraining. I think the best way is to provide training scripts as done by HRViT

MenghaoGuo commented

For VAN, it only uses 224 resolution for pretraining, which is same with MSCAN.