import torch from torch import nn from torch.nn import functional as F import pdb class CrossLayerFuse(nn.Module): def __init__(self, in_dims1, in_dims2, out_dims): super(CrossLayerFuse, self).__init__() self.linear = nn.Linear(in_dims1 + in_dims2, out_dims) self.adpool = nn.AdaptiveAvgPool2d((1, 1)) def forward(self, defea, x): x_pre = defea x = self.adpool(x).view(x.shape[0], x.shape[1]) x1 = torch.cat([x_pre, x], dim=1) x1 = self.linear(x1) return x1 class Transformer_Fusion(nn.Module): def __init__(self, dim=768, nhead=8, num_layers=1): super(Transformer_Fusion, self).__init__() self.decoder_layer = nn.TransformerDecoderLayer(d_model=dim, nhead=nhead) self.transformer_model = nn.TransformerDecoder(self.decoder_layer, num_layers=num_layers) def forward(self, vis, lan_full): WW, HH = vis.shape[2], vis.shape[3] vis = vis.view(vis.shape[0], vis.shape[1], -1) vis = vis.permute(2, 0, 1) lan = lan_full.permute(2, 0, 1) vis = self.transformer_model(vis, lan) vis = vis.permute(1, 2, 0) vis = vis.view(vis.shape[0], vis.shape[1], WW, HH) return vis class Language_Transformer(nn.Module): def __init__(self, hidden_size, lan_size): super(Language_Transformer, self).__init__() self.decoder_layer = nn.TransformerDecoderLayer(d_model=768, nhead=8) self.transformer_model = nn.TransformerDecoder(self.decoder_layer, num_layers=1) self.conv1 = nn.Conv2d(hidden_size, lan_size, 3, padding=1, bias=False) self.bn1 = nn.BatchNorm2d(lan_size) self.relu1 = nn.ReLU() def forward(self, vis, lan): vis = self.conv1(vis) vis = self.bn1(vis) vis = self.relu1(vis) vis = vis.view(vis.shape[0], vis.shape[1], -1) vis = vis.permute(2, 0, 1) lan = lan.permute(2, 0, 1) out = self.transformer_model(lan, vis) out = out.permute(1, 2, 0) return out class Decoder(nn.Module): def __init__(self, c4_dims, factor=2): super(Decoder, self).__init__() lan_size = 768 hidden_size = lan_size c4_size = c4_dims c3_size = c4_dims//(factor**1) c2_size = c4_dims//(factor**2) c1_size = c4_dims//(factor**3) self.adpool = nn.AdaptiveAvgPool2d((1, 1)) self.conv1_4 = nn.Conv2d(c4_size+c3_size, hidden_size, 3, padding=1, bias=False) self.bn1_4 = nn.BatchNorm2d(hidden_size) self.relu1_4 = nn.ReLU() self.conv2_4 = nn.Conv2d(hidden_size, hidden_size, 3, padding=1, bias=False) self.bn2_4 = nn.BatchNorm2d(hidden_size) self.relu2_4 = nn.ReLU() self.transformer_fusion1 = Transformer_Fusion(dim=768, nhead=8, num_layers=1) self.conv1_3 = nn.Conv2d(hidden_size + c2_size, hidden_size, 3, padding=1, bias=False) self.bn1_3 = nn.BatchNorm2d(hidden_size) self.relu1_3 = nn.ReLU() self.conv2_3 = nn.Conv2d(hidden_size, hidden_size, 3, padding=1, bias=False) self.bn2_3 = nn.BatchNorm2d(hidden_size) self.relu2_3 = nn.ReLU() self.crossfuse1 = CrossLayerFuse(hidden_size, hidden_size, lan_size) self.transformer_fusion2 = Transformer_Fusion(dim=768, nhead=8, num_layers=1) self.conv1_2 = nn.Conv2d(hidden_size + c1_size, hidden_size, 3, padding=1, bias=False) self.bn1_2 = nn.BatchNorm2d(hidden_size) self.relu1_2 = nn.ReLU() self.conv2_2 = nn.Conv2d(hidden_size, hidden_size, 3, padding=1, bias=False) self.bn2_2 = nn.BatchNorm2d(hidden_size) self.relu2_2 = nn.ReLU() self.conv1_1 = nn.Conv2d(hidden_size, 2, 1) self.lan_func = Language_Transformer(hidden_size, lan_size=768) self.crossfuse2 = CrossLayerFuse(lan_size, hidden_size, lan_size) def forward(self, lan_full, lan, x_c4, x_c3, x_c2, x_c1): # fuse Y4 and Y3 if x_c4.size(-2) < x_c3.size(-2) or x_c4.size(-1) < x_c3.size(-1): x_c4 = F.interpolate(input=x_c4, size=(x_c3.size(-2), x_c3.size(-1)), mode='bilinear', align_corners=True) x = torch.cat([x_c4, x_c3], dim=1) x = self.conv1_4(x) x = self.bn1_4(x) x = self.relu1_4(x) x = self.conv2_4(x) x = self.bn2_4(x) x = self.relu2_4(x) # [B, 512, 30, 30] de_feat = self.adpool(x).view(x.shape[0], x.shape[1]) x = self.transformer_fusion1(x, lan_full) # fuse top-down features and Y2 features and pre1 if x.size(-2) < x_c2.size(-2) or x.size(-1) < x_c2.size(-1): x = F.interpolate(input=x, size=(x_c2.size(-2), x_c2.size(-1)), mode='bilinear', align_corners=True) x = torch.cat([x, x_c2], dim=1) x = self.conv1_3(x) x = self.bn1_3(x) x = self.relu1_3(x) x = self.conv2_3(x) x = self.bn2_3(x) x = self.relu2_3(x) # [B, 512, 60, 60] new_lan = self.lan_func(x, lan) de_feat = self.crossfuse1(de_feat, x) x = self.transformer_fusion2(x, lan_full) # fuse top-down features and Y1 features if x.size(-2) < x_c1.size(-2) or x.size(-1) < x_c1.size(-1): x = F.interpolate(input=x, size=(x_c1.size(-2), x_c1.size(-1)), mode='bilinear', align_corners=True) x = torch.cat([x, x_c1], dim=1) x = self.conv1_2(x) x = self.bn1_2(x) x = self.relu1_2(x) x = self.conv2_2(x) x = self.bn2_2(x) x = self.relu2_2(x) # [B, 512, 120, 120] de_feat = self.crossfuse2(de_feat, x) return de_feat, new_lan, self.conv1_1(x)