Recent Progress/目前进度

[CWanli]

统一回复一下: 我最近沉迷股市，然后被市场教育了，现在打算回归初心，打工赚钱。代码很快就会开源，同时我也征求到了DMNet和APCNet作者Junjun He的同意，届时也会上传他的代码。各种细节也可以email我直接获取。

Dear All: I was addicted to stock market but I lose a lot of money. Now I come back to research & coding. The source code will release soon. Additionally, I was approved by Junjun He (the 1st author of DMNet and APCNet) to release his code. If you have any question, feel free to contact with me.

[ReckonerInheritor]

期待作者早日开源

[colorblank]

This is my version of TGM and TRM. But it could not work in segmentation task. I hope you can tell me where went wrong.

import torch
import torch.nn as nn
import torch.nn.functional as F

class LowRankModule(nn.Module):
    def __init__(self,in_channels, rank=8):
        super(LowRankModule, self).__init__()
        self.in_channels = in_channels
        if rank is None:
            self.rank = in_channels
        else:
            self.rank = rank
        self.lamd = nn.Parameter(torch.rand(self.rank)).cuda()
        self.conv_h_list = []
        self.conv_w_list = []
        self.conv_c_list = []
        for r in range(self.rank):
            self.conv_h_list.append(nn.Conv2d(1,1,1))
            self.conv_w_list.append(nn.Conv2d(1,1,1))
            self.conv_c_list.append(nn.Conv2d(in_channels,in_channels,1))
        self.conv_c = nn.ModuleList(self.conv_c_list)
        self.conv_h = nn.ModuleList(self.conv_h_list)
        self.conv_w = nn.ModuleList(self.conv_w_list)

        # self.conv_cat = nn.Conv2d(in_channels *2, in_channels,1)

    def forward(self,x):
        B,C,H,W = x.size()
        x_hp = F.adaptive_avg_pool3d(x, (1,H,1))
        x_wp = F.adaptive_avg_pool3d(x, (1,1,W))
        x_cp = F.adaptive_avg_pool3d(x,(C,1,1))

        attn_w = torch.zeros_like(x).to(x.device)
        for r in range(self.rank):
            x_c = self.conv_c[r](x_cp)  # [b,c,1,1]

            x_c = torch.sigmoid(x_c)
            x_h = self.conv_h[r](x_hp)  # [b,1,h,1]
            x_h = torch.sigmoid(x_h)
            x_w = self.conv_w[r](x_wp)  # [b,1,1,w]
            x_w = torch.sigmoid(x_w)

            x_c = x_c.view(B,C,-1)
            x_h = x_h.view(B,1,-1)
            x_tmp = torch.bmm(x_c,x_h)  # [b,c,h]
            x_tmp = x_tmp.view(B,C*H).unsqueeze(-1) # [b,c,h,1]

            x_w = x_w.view(B,1,-1)  # [b,1,w]
            x_final = torch.bmm(x_tmp,x_w)
            x_final = x_final.view(B,C,H, W)

            attn_w += self.lamd[r] * x_final

        out = x * attn_w

        return out

[CWanli]

This is my version of TGM and TRM. But it could not work in segmentation task. I hope you can tell me where went wrong.

import torch
import torch.nn as nn
import torch.nn.functional as F

class LowRankModule(nn.Module):
    def __init__(self,in_channels, rank=8):
        super(LowRankModule, self).__init__()
        self.in_channels = in_channels
        if rank is None:
            self.rank = in_channels
        else:
            self.rank = rank
        self.lamd = nn.Parameter(torch.rand(self.rank)).cuda()
        self.conv_h_list = []
        self.conv_w_list = []
        self.conv_c_list = []
        for r in range(self.rank):
            self.conv_h_list.append(nn.Conv2d(1,1,1))
            self.conv_w_list.append(nn.Conv2d(1,1,1))
            self.conv_c_list.append(nn.Conv2d(in_channels,in_channels,1))
        self.conv_c = nn.ModuleList(self.conv_c_list)
        self.conv_h = nn.ModuleList(self.conv_h_list)
        self.conv_w = nn.ModuleList(self.conv_w_list)

        # self.conv_cat = nn.Conv2d(in_channels *2, in_channels,1)

    def forward(self,x):
        B,C,H,W = x.size()
        x_hp = F.adaptive_avg_pool3d(x, (1,H,1))
        x_wp = F.adaptive_avg_pool3d(x, (1,1,W))
        x_cp = F.adaptive_avg_pool3d(x,(C,1,1))

        attn_w = torch.zeros_like(x).to(x.device)
        for r in range(self.rank):
            x_c = self.conv_c[r](x_cp)  # [b,c,1,1]

            x_c = torch.sigmoid(x_c)
            x_h = self.conv_h[r](x_hp)  # [b,1,h,1]
            x_h = torch.sigmoid(x_h)
            x_w = self.conv_w[r](x_wp)  # [b,1,1,w]
            x_w = torch.sigmoid(x_w)

            x_c = x_c.view(B,C,-1)
            x_h = x_h.view(B,1,-1)
            x_tmp = torch.bmm(x_c,x_h)  # [b,c,h]
            x_tmp = x_tmp.view(B,C*H).unsqueeze(-1) # [b,c,h,1]

            x_w = x_w.view(B,1,-1)  # [b,1,w]
            x_final = torch.bmm(x_tmp,x_w)
            x_final = x_final.view(B,C,H, W)

            attn_w += self.lamd[r] * x_final

        out = x * attn_w

        return out

Very close. I just uploaded my version.

[purse1996]

So the author codes require resolution of the feature map is 51288. While the convolution part of Colorblank's version is 1*1, which is different from the author's？