nn.PixelShuffle 与nn.Upsample的不同

f.interpolate和nn.upsample的区别

class MYGenerator(nn.Module): def __init__(self, n_residual=6): super().__init__() # 编码器（下采样 + 残差） self.down1 = nn.Sequential( nn.Conv2d(3, 64, 4, 2, 1), nn.InstanceNorm2d(64), nn.ReLU(inplace=True), CBAM(64), ResidualBlock(64) # 新增残差块 ) self.down2 = nn.Sequential( nn.Conv2d(64, 128, 4, 2, 1), nn.InstanceNorm2d(128), nn.ReLU(inplace=True), CBAM(128), ResidualBlock(128) # 新增残差块 ) self.down3 = nn.Sequential( nn.Conv2d(128, 256, 4, 2, 1), nn.InstanceNorm2d(256), nn.ReLU(inplace=True), CBAM(256), ResidualBlock(256) # 新增残差块 ) self.down4 = nn.Sequential( nn.Conv2d(256, 512, 4, 2, 1), nn.InstanceNorm2d(512), nn.ReLU(inplace=True), CBAM(512), ResidualBlock(512) # 新增残差块 ) # 瓶颈层（堆叠残差块） self.bottleneck = nn.Sequential( *[ResidualBlock(512) for _ in range(n_residual)] # 新增6个残差块 ) # 解码器（调整通道数适应拼接） self.up1 = nn.Sequential( nn.ConvTranspose2d(512, 256, 4, 2, 1), nn.InstanceNorm2d(256), nn.ReLU(inplace=True) ) self.up2 = nn.Sequential( nn.ConvTranspose2d(512, 128, 4, 2, 1), # 输入512=256(up1输出)+256(d3) nn.InstanceNorm2d(128), nn.ReLU(inplace=True) ) self.up3 = nn.Sequential( nn.ConvTranspose2d(256, 64, 4, 2, 1), # 输入256=128(up2输出)+128(d2) nn.InstanceNorm2d(64), nn.ReLU(inplace=True) ) self.up4 = nn.Sequential( nn.ConvTranspose2d(128, 3, 4, 2, 1), # 输入128=64(up3输出)+64(d1) nn.Tanh() ) def forward(self, x): # 编码 d1 = self.down1(x) # 64x128x128 d2 = self.down2(d1) # 128x64x64 d3 = self.down3(d2) # 256x32x32 d4 = self.down4(d3) # 512x16x16 # 瓶颈强化 d4 = self.bottleneck(d4) # 新增 # 解码（跳跃连接） u1 = self.up1(d4) # 256x32x32 u1 = torch.cat([u1, d3], dim=1) # 512x32x32 u2 = self.up2(u1) # 128x64x64 u2 = torch.cat([u2, d2], dim=1) # 256x64x64 u3 = self.up3(u2) # 64x128x128 u3 = torch.cat([u3, d1], dim=1) # 128x128x128 return self.up4(u3) # 3x256x256 该生成器应该如何让修改