Your work is really great! I would like to ask you about your EBGM module！

WCWCWCQAQAQA commented 6 days ago

Your improved attention module has deeply inspired my work, could you please share your EGSA code with me so that I can have a deeper understanding of your work?

YoeTon commented 2 days ago

你好，相关部分代码放在附件里了。

------------------ 原始邮件 ------------------ 发件人: "YoeTon/CTUNet" @.>; 发送时间: 2024年9月13日(星期五) 晚上6:43 @.>; @.***>; 主题: [YoeTon/CTUNet] Your work is really great! I would like to ask you about your EBGM module！ (Issue #1)

Your improved attention module has deeply inspired my work, could you please share your EGSA code with me so that I can have a deeper understanding of your work?

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread.Message ID: @.***>

WCWCWCQAQAQA commented 2 days ago

抱歉，我好像没有看到附件，您是不是不小心没点击上呀？

你好，相关部分代码放在附件里了。

------------------ 原始邮件 ------------------ 发件人: "YoeTon/CTUNet" @.>; 发送时间: 2024年9月13日(星期五) 晚上6:43 @.>; @.***>; 主题: [YoeTon/CTUNet] Your work is really great! I would like to ask you about your EBGM module！ (Issue #1)

Your improved attention module has deeply inspired my work, could you please share your EGSA code with me so that I can have a deeper understanding of your work?

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread.Message ID: @.***>

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

YoeTon commented 2 days ago

class EGSA(nn.Module):

def init(self, dim, num_heads, bias): super(EGSA, self).init() self.num_heads = num_heads self.temperature = nn.Parameter(torch.ones(num_heads, 1, 1)) self.conv0 = nn.Conv2d(dim, dim, kernel_size=1, bias=bias) self.edge_body = edge(dim) self.conv1 = nn.Conv2d(dim, dim, kernel_size=1, bias=bias) self.project_out = nn.Conv2d(dim, dim, kernel_size=1, bias=bias) self.relu = nn.ReLU(inplace=False)

def forward(self, x): b, c, h, w = x.shape x_edge = self.edge_body(x) x_q = x + 0.01 * x_edge q = self.conv0(x_q) k = q v = self.conv1(x)

q = rearrange(q, 'b (head c) h w -> b head c (h w)', head=self.num_heads) k = rearrange(k, 'b (head c) h w -> b head c (h w)', head=self.num_heads) v = rearrange(v, 'b (head c) h w -> b head c (h w)', head=self.num_heads) q = torch.nn.functional.normalize(q, dim=-1) k = torch.nn.functional.normalize(k, dim=-1)

attn = (q @ k.transpose(-2, -1)) * self.temperature attn = attn.softmax(dim=-1)

out = (attn @ v)

out = rearrange(out, 'b head c (h w) -> b (head c) h w', head=self.num_heads, h=h, w=w) out = self.project_out(out) return out

class edge(nn.Module): def init(self, n_feats): super(edge, self).init() self.conv = nn.Conv2d(n_feats * 2, 2, kernel_size=(3, 3), stride=1, padding=(1, 1))

def forward(self, out): size = out.size()[2:] flow = F.interpolate(out, (int(size[0] / 4.), int(size[1] / 4.)), mode='bilinear', align_corners=False) flow = F.interpolate(flow, size, mode='bilinear', align_corners=False) flow = self.conv(torch.cat([out, flow], 1)) flow = flow_warp(out, flow, size)

return out - flow

def flow_warp(input, flow, size): out_h, out_w = size

n, c, h, w = input.size()

norm = torch.tensor([[[[out_w, out_h]]]]).type_as(input).to(input.device)

h_grid = torch.linspace(-1.0, 1.0, out_h).view(-1, 1).repeat(1, out_w) w_gird = torch.linspace(-1.0, 1.0, out_w).repeat(out_h, 1) grid = torch.cat((w_gird.unsqueeze(2), h_grid.unsqueeze(2)), 2)

grid = grid.repeat(n, 1, 1, 1).type_as(input).to(input.device) grid = grid + flow.permute(0, 2, 3, 1) / norm

output = F.grid_sample(input, grid, align_corners=False)

return output

---Original--- From: "Rui @.> Date: Tue, Sep 17, 2024 18:09 PM To: @.>; Cc: @.**@.>; Subject: Re: [YoeTon/CTUNet] Your work is really great! I would like to ask you about your EBGM module！ (Issue #1)

抱歉，我好像没有看到附件，您是不是不小心没点击上呀？

你好，相关部分代码放在附件里了。

------------------ 原始邮件 ------------------ 发件人: "YoeTon/CTUNet" @.>; 发送时间: 2024年9月13日(星期五) 晚上6:43 @.>; @.***>; 主题: [YoeTon/CTUNet] Your work is really great! I would like to ask you about your EBGM module！ (Issue #1)

Your improved attention module has deeply inspired my work, could you please share your EGSA code with me so that I can have a deeper understanding of your work?

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread.Message ID: @.***>

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.> — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.>

WCWCWCQAQAQA commented 2 days ago

谢谢您的耐心回复！已看到！

YoeTon / CTUNet

Your work is really great! I would like to ask you about your EBGM module！ #1