import torch import torch.nn as nn from torchvision import models def convrelu(in_channels, out_channels, kernel, padding, pool): return nn.Sequential( nn.Conv2d(in_channels, out_channels, kernel, padding=padding), #In conv, the dimension of the output, if the input is H,W, is # H+2*padding-kernel +1 nn.ReLU(inplace=True), nn.MaxPool2d(pool, stride=pool, padding=0, dilation=1, return_indices=False, ceil_mode=False) #pooling takes Height H and width W to (H-pool)/pool+1 = H/pool, and floor. Same for W. #altogether, the output size is (H+2*padding-kernel +1)/pool. ) def convreluT(in_channels, out_channels, kernel, padding): return nn.Sequential( nn.ConvTranspose2d(in_channels, out_channels, kernel, stride=2, padding=padding), nn.ReLU(inplace=True) #input is H X W, output is (H-1)*2 - 2*padding + kernel ) class RadioWNet(nn.Module): def __init__(self,inputs=2,phase="firstU"): super().__init__() self.inputs=inputs self.phase=phase if inputs<=3: self.layer00 = convrelu(inputs, 6, 3, 1,1) self.layer0 = convrelu(6, 40, 5, 2,2) else: self.layer00 = convrelu(inputs, 10, 3, 1,1) self.layer0 = convrelu(10, 40, 5, 2,2) self.layer1 = convrelu(40, 50, 5, 2,2) self.layer10 = convrelu(50, 60, 5, 2,1) self.layer2 = convrelu(60, 100, 5, 2,2) self.layer20 = convrelu(100, 100, 3, 1,1) self.layer3 = convrelu(100, 150, 5, 2,2) self.layer4 =convrelu(150, 300, 5, 2,2) self.layer5 =convrelu(300, 500, 5, 2,2) self.conv_up5 =convreluT(500, 300, 4, 1) self.conv_up4 = convreluT(300+300, 150, 4, 1) self.conv_up3 = convreluT(150 + 150, 100, 4, 1) self.conv_up20 = convrelu(100 + 100, 100, 3, 1, 1) self.conv_up2 = convreluT(100 + 100, 60, 6, 2) self.conv_up10 = convrelu(60 + 60, 50, 5, 2, 1) self.conv_up1 = convreluT(50 + 50, 40, 6, 2) self.conv_up0 = convreluT(40 + 40, 20, 6, 2) if inputs<=3: self.conv_up00 = convrelu(20+6+inputs, 20, 5, 2,1) else: self.conv_up00 = convrelu(20+10+inputs, 20, 5, 2,1) self.conv_up000 = convrelu(20+inputs, 1, 5, 2,1) self.Wlayer00 = convrelu(inputs+1, 20, 3, 1,1) self.Wlayer0 = convrelu(20, 30, 5, 2,2) self.Wlayer1 = convrelu(30, 40, 5, 2,2) self.Wlayer10 = convrelu(40, 50, 5, 2,1) self.Wlayer2 = convrelu(50, 60, 5, 2,2) self.Wlayer20 = convrelu(60, 70, 3, 1,1) self.Wlayer3 = convrelu(70, 90, 5, 2,2) self.Wlayer4 =convrelu(90, 110, 5, 2,2) self.Wlayer5 =convrelu(110, 150, 5, 2,2) self.Wconv_up5 =convreluT(150, 110, 4, 1) self.Wconv_up4 = convreluT(110+110, 90, 4, 1) self.Wconv_up3 = convreluT(90 + 90, 70, 4, 1) self.Wconv_up20 = convrelu(70 + 70, 60, 3, 1, 1) self.Wconv_up2 = convreluT(60 + 60, 50, 6, 2) self.Wconv_up10 = convrelu(50 + 50, 40, 5, 2, 1) self.Wconv_up1 = convreluT(40 + 40, 30, 6, 2) self.Wconv_up0 = convreluT(30 + 30, 20, 6, 2) self.Wconv_up00 = convrelu(20+20+inputs+1, 20, 5, 2,1) self.Wconv_up000 = convrelu(20+inputs+1, 1, 5, 2,1) def forward(self, input): input0=input[:,0:self.inputs,:,:] if self.phase=="firstU": layer00 = self.layer00(input0) layer0 = self.layer0(layer00) layer1 = self.layer1(layer0) layer10 = self.layer10(layer1) layer2 = self.layer2(layer10) layer20 = self.layer20(layer2) layer3 = self.layer3(layer20) layer4 = self.layer4(layer3) layer5 = self.layer5(layer4) layer4u = self.conv_up5(layer5) layer4u = torch.cat([layer4u, layer4], dim=1) layer3u = self.conv_up4(layer4u) layer3u = torch.cat([layer3u, layer3], dim=1) layer20u = self.conv_up3(layer3u) layer20u = torch.cat([layer20u, layer20], dim=1) layer2u = self.conv_up20(layer20u) layer2u = torch.cat([layer2u, layer2], dim=1) layer10u = self.conv_up2(layer2u) layer10u = torch.cat([layer10u, layer10], dim=1) layer1u = self.conv_up10(layer10u) layer1u = torch.cat([layer1u, layer1], dim=1) layer0u = self.conv_up1(layer1u) layer0u = torch.cat([layer0u, layer0], dim=1) layer00u = self.conv_up0(layer0u) layer00u = torch.cat([layer00u, layer00], dim=1) layer00u = torch.cat([layer00u,input0], dim=1) layer000u = self.conv_up00(layer00u) layer000u = torch.cat([layer000u,input0], dim=1) output1 = self.conv_up000(layer000u) Winput=torch.cat([output1, input], dim=1).detach() Wlayer00 = self.Wlayer00(Winput).detach() Wlayer0 = self.Wlayer0(Wlayer00).detach() Wlayer1 = self.Wlayer1(Wlayer0).detach() Wlayer10 = self.Wlayer10(Wlayer1).detach() Wlayer2 = self.Wlayer2(Wlayer10).detach() Wlayer20 = self.Wlayer20(Wlayer2).detach() Wlayer3 = self.Wlayer3(Wlayer20).detach() Wlayer4 = self.Wlayer4(Wlayer3).detach() Wlayer5 = self.Wlayer5(Wlayer4).detach() Wlayer4u = self.Wconv_up5(Wlayer5).detach() Wlayer4u = torch.cat([Wlayer4u, Wlayer4], dim=1).detach() Wlayer3u = self.Wconv_up4(Wlayer4u).detach() Wlayer3u = torch.cat([Wlayer3u, Wlayer3], dim=1).detach() Wlayer20u = self.Wconv_up3(Wlayer3u).detach() Wlayer20u = torch.cat([Wlayer20u, Wlayer20], dim=1).detach() Wlayer2u = self.Wconv_up20(Wlayer20u).detach() Wlayer2u = torch.cat([Wlayer2u, Wlayer2], dim=1).detach() Wlayer10u = self.Wconv_up2(Wlayer2u).detach() Wlayer10u = torch.cat([Wlayer10u, Wlayer10], dim=1).detach() Wlayer1u = self.Wconv_up10(Wlayer10u).detach() Wlayer1u = torch.cat([Wlayer1u, Wlayer1], dim=1).detach() Wlayer0u = self.Wconv_up1(Wlayer1u).detach() Wlayer0u = torch.cat([Wlayer0u, Wlayer0], dim=1).detach() Wlayer00u = self.Wconv_up0(Wlayer0u).detach() Wlayer00u = torch.cat([Wlayer00u, Wlayer00], dim=1).detach() Wlayer00u = torch.cat([Wlayer00u,Winput], dim=1).detach() Wlayer000u = self.Wconv_up00(Wlayer00u).detach() Wlayer000u = torch.cat([Wlayer000u,Winput], dim=1).detach() output2 = self.Wconv_up000(Wlayer000u).detach() else: layer00 = self.layer00(input0).detach() layer0 = self.layer0(layer00).detach() layer1 = self.layer1(layer0).detach() layer10 = self.layer10(layer1).detach() layer2 = self.layer2(layer10).detach() layer20 = self.layer20(layer2).detach() layer3 = self.layer3(layer20).detach() layer4 = self.layer4(layer3).detach() layer5 = self.layer5(layer4).detach() layer4u = self.conv_up5(layer5).detach() layer4u = torch.cat([layer4u, layer4], dim=1).detach() layer3u = self.conv_up4(layer4u).detach() layer3u = torch.cat([layer3u, layer3], dim=1).detach() layer20u = self.conv_up3(layer3u).detach() layer20u = torch.cat([layer20u, layer20], dim=1).detach() layer2u = self.conv_up20(layer20u).detach() layer2u = torch.cat([layer2u, layer2], dim=1).detach() layer10u = self.conv_up2(layer2u).detach() layer10u = torch.cat([layer10u, layer10], dim=1).detach() layer1u = self.conv_up10(layer10u).detach() layer1u = torch.cat([layer1u, layer1], dim=1).detach() layer0u = self.conv_up1(layer1u).detach() layer0u = torch.cat([layer0u, layer0], dim=1).detach() layer00u = self.conv_up0(layer0u).detach() layer00u = torch.cat([layer00u, layer00], dim=1).detach() layer00u = torch.cat([layer00u,input0], dim=1).detach() layer000u = self.conv_up00(layer00u).detach() layer000u = torch.cat([layer000u,input0], dim=1).detach() output1 = self.conv_up000(layer000u).detach() Winput=torch.cat([output1, input], dim=1).detach() Wlayer00 = self.Wlayer00(Winput) Wlayer0 = self.Wlayer0(Wlayer00) Wlayer1 = self.Wlayer1(Wlayer0) Wlayer10 = self.Wlayer10(Wlayer1) Wlayer2 = self.Wlayer2(Wlayer10) Wlayer20 = self.Wlayer20(Wlayer2) Wlayer3 = self.Wlayer3(Wlayer20) Wlayer4 = self.Wlayer4(Wlayer3) Wlayer5 = self.Wlayer5(Wlayer4) Wlayer4u = self.Wconv_up5(Wlayer5) Wlayer4u = torch.cat([Wlayer4u, Wlayer4], dim=1) Wlayer3u = self.Wconv_up4(Wlayer4u) Wlayer3u = torch.cat([Wlayer3u, Wlayer3], dim=1) Wlayer20u = self.Wconv_up3(Wlayer3u) Wlayer20u = torch.cat([Wlayer20u, Wlayer20], dim=1) Wlayer2u = self.Wconv_up20(Wlayer20u) Wlayer2u = torch.cat([Wlayer2u, Wlayer2], dim=1) Wlayer10u = self.Wconv_up2(Wlayer2u) Wlayer10u = torch.cat([Wlayer10u, Wlayer10], dim=1) Wlayer1u = self.Wconv_up10(Wlayer10u) Wlayer1u = torch.cat([Wlayer1u, Wlayer1], dim=1) Wlayer0u = self.Wconv_up1(Wlayer1u) Wlayer0u = torch.cat([Wlayer0u, Wlayer0], dim=1) Wlayer00u = self.Wconv_up0(Wlayer0u) Wlayer00u = torch.cat([Wlayer00u, Wlayer00], dim=1) Wlayer00u = torch.cat([Wlayer00u,Winput], dim=1) Wlayer000u = self.Wconv_up00(Wlayer00u) Wlayer000u = torch.cat([Wlayer000u,Winput], dim=1) output2 = self.Wconv_up000(Wlayer000u) return [output1,output2]