unet up block

unet_up

# a sample up block
def make_conv_bn_relu(in_channels, out_channels, kernel_size=3, stride=1, padding=1):
    return [
        nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size,  stride=stride, padding=padding, bias=False),
        nn.BatchNorm2d(out_channels),
        nn.ReLU(inplace=True)
    ]
self.up4 = nn.Sequential(
    *make_conv_bn_relu(128,64, kernel_size=3, stride=1, padding=1 ),
    *make_conv_bn_relu(64,64, kernel_size=3, stride=1, padding=1 )
)
self.final_conv = nn.Conv2d(32, num_classes, kernel_size=1, stride=1, padding=0 )

# upsample out_last, concatenate with down1 and apply conv operations
out   = F.upsample(out_last, scale_factor=2, mode='bilinear')  
out   = torch.cat([down1, out], 1)
out   = self.up4(out)

# final 1x1 conv for predictions
final_out = self.final_conv(out)