import torch from torch import nn from torch.nn import functional as F from einops import rearrange, repeat import math def scaled_dot_product(q, k, v, mask=None): d_k = q.size()[-1] attn_logits = torch.matmul(q, k.transpose(-2, -1)) attn_logits = attn_logits / math.sqrt(d_k) if mask is not None: attn_logits = attn_logits.masked_fill(mask == 0, -9e15) attention = F.softmax(attn_logits, dim=-1) values = torch.matmul(attention, v) return values, attention """ Fusion Module""" class ASM(nn.Module): def __init__(self, in_channels, all_channels): super(ASM, self).__init__() self.non_local = NonLocalBlock(in_channels) def forward(self, lc, fuse, gc): fuse = self.non_local(fuse) fuse = torch.cat([lc, fuse, gc], dim=1) return fuse """ Squeeze and Excitation Layer https://arxiv.org/abs/1709.01507 """ class SELayer(nn.Module): def __init__(self, channel, reduction=16): super(SELayer, self).__init__() self.avg_pool = nn.AdaptiveAvgPool2d(1) self.fc = nn.Sequential( nn.Linear(channel, channel // reduction, bias=False), nn.ReLU(inplace=True), nn.Linear(channel // reduction, channel, bias=False), nn.Sigmoid() ) def forward(self, x): b, c, _, _ = x.size() y = self.avg_pool(x).view(b, c) y = self.fc(y).view(b, c, 1, 1) return x * y.expand_as(x) """ Non Local Block https://arxiv.org/abs/1711.07971 """ class NonLocalBlock(nn.Module): def __init__(self, in_channels, inter_channels=None, sub_sample=True, bn_layer=True): super(NonLocalBlock, self).__init__() self.sub_sample = sub_sample self.in_channels = in_channels self.inter_channels = inter_channels if self.inter_channels is None: self.inter_channels = in_channels // 2 if self.inter_channels == 0: self.inter_channels = 1 self.g = nn.Conv2d(in_channels=self.in_channels, out_channels=self.inter_channels, kernel_size=1, stride=1, padding=0) if bn_layer: self.W = nn.Sequential( nn.Conv2d(in_channels=self.inter_channels, out_channels=self.in_channels, kernel_size=1, stride=1, padding=0), nn.BatchNorm2d(self.in_channels) ) nn.init.constant_(self.W[1].weight, 0) nn.init.constant_(self.W[1].bias, 0) else: self.W = nn.Conv2d(in_channels=self.inter_channels, out_channels=self.in_channels, kernel_size=1, stride=1, padding=0) nn.init.constant_(self.W.weight, 0) nn.init.constant_(self.W.bias, 0) self.theta = nn.Conv2d(in_channels=self.in_channels, out_channels=self.inter_channels, kernel_size=1, stride=1, padding=0) self.phi = nn.Conv2d(in_channels=self.in_channels, out_channels=self.inter_channels, kernel_size=1, stride=1, padding=0) if sub_sample: self.g = nn.Sequential(self.g, nn.MaxPool2d(kernel_size=(2, 2))) self.phi = nn.Sequential(self.phi, nn.MaxPool2d(kernel_size=(2, 2))) def forward(self, x): batch_size = x.size(0) g_x = self.g(x).view(batch_size, self.inter_channels, -1) g_x = g_x.permute(0, 2, 1) theta_x = self.theta(x).view(batch_size, self.inter_channels, -1) theta_x = theta_x.permute(0, 2, 1) phi_x = self.phi(x).view(batch_size, self.inter_channels, -1) f = torch.matmul(theta_x, phi_x) f_div_C = F.softmax(f, dim=-1) y = torch.matmul(f_div_C, g_x) y = y.permute(0, 2, 1).contiguous() y = y.view(batch_size, self.inter_channels, *x.size()[2:]) W_y = self.W(y) z = W_y + x return z #AGCM Module class CrossNonLocalBlock(nn.Module): def __init__(self, in_channels_source,in_channels_target, inter_channels, sub_sample=False, bn_layer=True): super(CrossNonLocalBlock, self).__init__() self.sub_sample = sub_sample self.in_channels_source = in_channels_source self.in_channels_target = in_channels_target self.inter_channels = inter_channels """ if self.inter_channels is None: self.inter_channels = in_channels // 2 if self.inter_channels == 0: self.inter_channels = 1 """ self.g = nn.Conv2d(in_channels=self.in_channels_source, out_channels=self.inter_channels, kernel_size=1, stride=1, padding=0) self.theta = nn.Conv2d(in_channels=self.in_channels_source, out_channels=self.inter_channels, kernel_size=1, stride=1, padding=0) self.phi = nn.Conv2d(in_channels=self.in_channels_target, out_channels=self.inter_channels, kernel_size=1, stride=1, padding=0) if bn_layer: self.W = nn.Sequential( nn.Conv2d(in_channels=self.inter_channels, out_channels=self.in_channels_target, kernel_size=1, stride=1, padding=0), nn.BatchNorm2d(self.in_channels_target) ) nn.init.constant_(self.W[1].weight, 0) nn.init.constant_(self.W[1].bias, 0) if sub_sample: self.g = nn.Sequential(self.g, nn.MaxPool2d(kernel_size=(2, 2))) self.phi = nn.Sequential(self.phi, nn.MaxPool2d(kernel_size=(2, 2))) def forward(self,x,l): batch_size = x.size(0) g_x = self.g(x).view(batch_size, self.inter_channels, -1) g_x = g_x.permute(0, 2, 1) #source theta_x1 = self.theta(x) theta_x = self.theta(x).view(batch_size, self.inter_channels, -1) theta_x = theta_x.permute(0, 2, 1) #source phi_x = self.phi(l).view(batch_size, self.inter_channels, -1) #target f = torch.matmul(theta_x, phi_x) f_div_C = F.softmax(f, dim=-1) f_div_C = f_div_C.permute(0,2,1) y = torch.matmul(f_div_C, g_x) y = y.permute(0, 2, 1).contiguous() y = y.view(batch_size, self.inter_channels, *l.size()[2:]) W_y = self.W(y) z = W_y + l return z #SFEM module class NonLocalBlock_PatchWise(nn.Module): def __init__(self, in_channel, inter_channel, patch_factor): super(NonLocalBlock_PatchWise, self).__init__() "Embedding dimension must be 0 modulo number of heads." self.in_channel = in_channel self.patch_factor = patch_factor self.patch_width = int(8/self.patch_factor) self.patch_height = int(8/self.patch_factor) self.stride_width = int(8/self.patch_factor) self.stride_height = int(8/self.patch_factor) self.unfold = nn.Unfold(kernel_size=(self.patch_width, self.patch_height), stride=(self.stride_width, self.stride_height)) self.adp = nn.AdaptiveAvgPool2d(8) self.bottleneck = nn.Conv2d(64,inter_channel,kernel_size=(1,1)) self.non_block = NonLocalBlock(self.in_channel) self.adp_post = nn.AdaptiveAvgPool2d((8,8)) def forward(self, x): batch_size = x.size(0) x_up = self.adp(x) x_up = self.unfold(x) batch_size,p_dim,p_size = x_up.size() x_up = x_up.view(batch_size,-1,self.in_channel,p_size) final_output = torch.tensor([]).cuda() index = torch.range(0,p_size,1,dtype=torch.int64).cuda() for i in range(int(p_size)): divide = torch.index_select(x_up, 3, index[i]) divide = divide.view(batch_size,-1,self.in_channel) patch_width = int(divide.size(1) ** 0.5) divide = divide.reshape(batch_size,self.in_channel,patch_width,patch_width) # tensor to operate on attn = self.non_block(divide) output = attn.view(batch_size,-1,self.in_channel,1) final_output = torch.cat((final_output,output),dim=3) final_output = final_output.view(batch_size, self.in_channel, 8,8) return final_output class GCM_up(nn.Module): def __init__(self, in_channels, out_channels): super(GCM_up, self).__init__() self.adp = nn.AdaptiveAvgPool2d((8,8)) self.patch1 = NonLocalBlock_PatchWise(in_channels,out_channels,2) self.patch2 = NonLocalBlock_PatchWise(in_channels,out_channels,4) self.patch3 = NonLocalBlock(256,64) self.fuse = SELayer(3*256) self.conv = nn.Conv2d(3*256, out_channels, 1, 1) self.relu = nn.ReLU(inplace=True) def forward(self, x): b,c,h,w = x.size() x = self.adp(x) patch1 = self.patch1(x) patch2 = self.patch2(x) patch3 = self.patch3(x) global_cat = torch.cat((patch1, patch2, patch3), dim=1) fuse = self.relu(self.conv(self.fuse(global_cat))) adp_post = nn.AdaptiveAvgPool2d((h,w)) fuse = adp_post(fuse) return fuse