search

Visual-Attention-Network / SegNeXt

Official Pytorch implementations for "SegNeXt: Rethinking Convolutional Attention Design for Semantic Segmentation" (NeurIPS 2022)
Apache License 2.0
796 stars 85 forks source link

MSCAN-Large's e.r or mlp_ratios needs to fix #2

Closed saicoco closed 2 years ago

saicoco commented 2 years ago

https://github.com/Visual-Attention-Network/SegNeXt/blob/1e51c8ac5b94907d915346d5c0e998093d97aa93/mmseg/models/backbones/mscan.py#L175

It will raise size mismatch when load mscan-large reference mscan-large-config

saicoco commented 2 years ago

mlp_raitos=[8, 8, 4, 4]

uyzhang commented 2 years ago

We write the mlp_raitos parameter in https://github.com/Visual-Attention-Network/SegNeXt/blob/main/local_configs/_base_/models/mscan.py#L10

saicoco commented 2 years ago

Thanks to reply, I just miss it