[FIX] Bug fix

networks.py

@@ -26,6 +26,22 @@ def forward(self, inp):
 class PanNet(nn.Module):
     def __init__(self, nbands, ratio):
         super(PanNet, self).__init__()
+
+        bfilter_ms = torch.ones(nbands, 1, 5, 5)
+        bfilter_ms = bfilter_ms / (bfilter_ms.shape[-2] * bfilter_ms.shape[-1])
+        bfilter_pan = torch.ones(1, 1, 5, 5)
+        bfilter_pan = bfilter_pan / (bfilter_pan.shape[-2] * bfilter_pan.shape[-1])
+
+        self.dephtconv_ms = nn.Conv2d(in_channels=nbands, out_channels=nbands, padding=(2, 2),
+                         kernel_size=bfilter_ms.shape, groups=nbands, bias=False, padding_mode='replicate')
+        self.dephtconv_ms.weight.data = bfilter_ms
+        self.dephtconv_ms.weight.requires_grad = False
+
+        self.dephtconv_pan = nn.Conv2d(in_channels=1, out_channels=1, padding=(2, 2),
+                                      kernel_size=bfilter_pan.shape, groups=1, bias=False, padding_mode='replicate')
+        self.dephtconv_pan.weight.data = bfilter_pan
+        self.dephtconv_pan.weight.requires_grad = False
+
         self.ratio = ratio
         self.Conv2d_transpose = nn.ConvTranspose2d(nbands, nbands, 8, 4, padding=(2, 2), bias=False)
         self.Conv = nn.Conv2d(nbands + 1, 32, 3, padding=(1, 1))
@@ -43,10 +59,13 @@ def forward(self, inp):
         lms = inp[:, :-1, 2::self.ratio, 2::self.ratio]
         pan = torch.unsqueeze(inp[:, -1, :, :], dim=1)
 
-        x = self.Conv2d_transpose(lms)
-        x = torch.cat((x, pan), dim=1)
+        lms_hp = lms - self.dephtconv_ms(lms)
+        pan_hp = pan - self.dephtconv_pan(pan)
 
-        x1 = F.relu(self.Conv(x))
+        x = self.Conv2d_transpose(lms_hp)
+        net_inp = torch.cat((x, pan_hp), dim=1)
+
+        x1 = F.relu(self.Conv(net_inp))
 
         x2 = F.relu(self.Conv_1(x1))
         x3 = self.Conv_2(x2) + x1
@@ -62,7 +81,9 @@ def forward(self, inp):
 
         x10 = self.Conv_9(x9)
 
-        return x10
+        x11 = inp[:, :-1, :, :] + x10
+
+        return x11
 
 
 class DRPNN(nn.Module):