Created by: shengkelong

    #MSRB
    class MSRB(fluid.dygraph.Layer):                                                                                                                                                                                                                    
def __init__(self,channels=64):
    super(MSRB, self).__init__()
    self.conv_3_1 = Conv2D (channels,channels,3,padding = 1,act='relu',param_attr=fluid.ParamAttr(initializer=fluid.initializer.NormalInitializer(scale=0.001)))
    self.conv_3_2 = Conv2D (channels*2,channels*2,3,padding = 1,act='relu',param_attr=fluid.ParamAttr(initializer=fluid.initializer.NormalInitializer(scale=0.001)))
    self.conv_5_1 = Conv2D (channels,channels,5,padding = 2,act='relu',param_attr=fluid.ParamAttr(initializer=fluid.initializer.NormalInitializer(scale=0.001)))
    self.conv_5_2 = Conv2D (channels*2,channels*2,5,padding = 2,act='relu',param_attr=fluid.ParamAttr(initializer=fluid.initializer.NormalInitializer(scale=0.001)))
    self.confusion = Conv2D (channels*4,channels,1,padding = 0)
def forward(self, x):
    input_1 = x
    output_3_1 = self.conv_3_1(input_1)
    output_5_1 = self.conv_5_1(input_1)
    input_2 = fluid.layers.concat([output_3_1, output_5_1], axis=1)
    output_3_2 = self.conv_3_2(input_2)
    output_5_2 = self.conv_5_2(input_2)
    input_3 = fluid.layers.concat([output_3_2, output_5_2], axis=1)
    output = self.confusion(input_3)
    output += x
    return output
   
    
    #MSRN
    class MSRN(fluid.dygraph.Layer):
def __init__(self):
   
    # define head module
    super(MSRN, self).__init__()
    self.n_blocks = 4
    
    
    modules_head = [Conv2D(3, 64, 3,padding = 1)]

    # define body module
    modules_body = paddle.fluid.dygraph.LayerList([MSRB()for i in range(self.n_blocks)])
    

    # define tail module
    modules_tail = [
        Conv2D(64 * (self.n_blocks + 1), 64, 1),
        Conv2D(64, 64, 3,padding = 1),
        upsample(),
        Conv2D(64, 3, 3,padding = 1)]
    self.head = fluid.dygraph.Sequential(*modules_head)
    self.body = fluid.dygraph.Sequential(*modules_body)
    self.tail = fluid.dygraph.Sequential(*modules_tail)

def forward(self, x):
    x = self.head(x)
    res = x

    MSRB_out = []
    for i in range(self.n_blocks):
        x = self.body[i](x)
        MSRB_out.append(x)
    MSRB_out.append(res)

    res = fluid.layers.concat(MSRB_out,axis=1)
    x = self.tail(res)
    return x

复现MSRN超分辨网络，效果不佳，psnr只能到35，一旦使用默认初始化损失会直逼1e10很难收敛，不知是什么原因