TopFormer implementation differs from original reference implementation

[dominikandreas]

sig_act is computed differently from the original implementation. Complare

PaddleViT/semantic_segmentation/src/models/backbones/top_transformer.py

Lines 330 to 334 in 55b33c3

	global_act = self.global_act(x_global)
	global_act = self.act(global_act)
	global_act = nn.functional.interpolate(
	global_act, size=(H, W), mode='bilinear', align_corners=False)
	sigmoid_act = F.interpolate(self.act(global_act), size=(H, W), mode='bilinear', align_corners=False)

with

https://github.com/hustvl/TopFormer/blob/2dc253c49ef78742ca6b44e550c5fea63a274288/mmseg/models/backbones/topformer.py#L328

I assume this is not intentional. The fix is straightforward:

    def forward(self, x_local, x_global):
        '''
        x_g: global features
        x_l: local features
        '''
        B, C, H, W = x_local.shape
        local_feat = self.local_embedding(x_local)
        
        global_act = self.global_act(x_global)
        sig_act = F.interpolate(self.act(global_act), size=(H, W), mode='bilinear', align_corners=False)
        
        global_feat = self.global_embedding(x_global)
        global_feat = F.interpolate(global_feat, size=(H, W), mode='bilinear', align_corners=False)
        
        out = local_feat * sig_act + global_feat
        return out