Created by: Costwen

`

class adaILN(fluid.dygraph.Layer):

def __init__(self, in_channels, eps=1e-5):
    super(adaILN, self).__init__()
    self.eps = eps
    self.rho = fluid.layers.create_parameter((1, in_channels, 1, 1), dtype='float32', is_bias=True,
    default_initializer=fluid.initializer.ConstantInitializer(0.9))

def var(self, input, dim):
    mean = fluid.layers.reduce_mean(input, dim, keep_dim=True)
    tmp = fluid.layers.reduce_mean((input - mean)**2, dim, keep_dim=True)
    return tmp

def forward(self, input, gamma, beta):
    in_mean, in_var = fluid.layers.reduce_mean(input, dim=[2,3], keep_dim=True), self.var(input, dim =[2,3])
    ln_mean, ln_var = fluid.layers.reduce_mean(input, dim=[1,2,3], keep_dim=True), self.var(input, dim=[1,2,3])
    out_in = (input - in_mean) / fluid.layers.sqrt(in_var + self.eps)
    out_ln = (input - ln_mean) / fluid.layers.sqrt(ln_var + self.eps)
    ex_rho = fluid.layers.expand(self.rho, expand_times = [input.shape[0], 1, 1, 1])
    print("rho",fluid.layers.reduce_mean(self.rho).numpy())
    out = ex_rho * out_in + (1 - ex_rho)*out_ln
    out = out * gamma + beta
    return out

class ILN(fluid.dygraph.Layer):

def __init__(self, in_channels, eps=1e-5):
    super(ILN, self).__init__()
    self.eps = eps
    self.rho = fluid.layers.create_parameter((1, in_channels, 1, 1), dtype='float32', is_bias=True,
    default_initializer=fluid.initializer.ConstantInitializer(0.0))
    self.gamma = fluid.layers.create_parameter((1, in_channels, 1, 1), dtype='float32', is_bias=True,
    default_initializer=fluid.initializer.ConstantInitializer(1.0))
    self.beta = fluid.layers.create_parameter((1, in_channels, 1, 1), dtype='float32', is_bias=True,
    default_initializer=fluid.initializer.ConstantInitializer(0.0))
    
def var(self, input, dim):
    mean = fluid.layers.reduce_mean(input, dim, keep_dim=True)
    tmp = fluid.layers.reduce_mean((input - mean)**2, dim, keep_dim=True)
    return tmp

def forward(self, input):
    in_mean, in_var = fluid.layers.reduce_mean(input, dim=[2,3], keep_dim=True), self.var(input, dim =[2,3])
    ln_mean, ln_var = fluid.layers.reduce_mean(input, dim=[1,2,3], keep_dim=True), self.var(input, dim=[1,2,3])
    out_in = (input - in_mean) / fluid.layers.sqrt(in_var + self.eps)
    out_ln = (input - ln_mean) / fluid.layers.sqrt(ln_var + self.eps)
    ex_rho = fluid.layers.expand(self.rho, expand_times = [input.shape[0], 1, 1, 1])
    ex_gamma = fluid.layers.expand(self.gamma, expand_times = [input.shape[0], 1, 1, 1])
    ex_beta = fluid.layers.expand(self.beta, expand_times = [input.shape[0], 1, 1, 1])
    print("rho",fluid.layers.reduce_mean(self.rho).numpy())
    print("gamma",fluid.layers.reduce_mean(self.gamma).numpy())
    print("beta",fluid.layers.reduce_mean(self.beta).numpy())
    out = ex_rho * out_in + (1 - ex_rho) * out_ln
    out = out * ex_gamma + ex_beta
    return out

` 通过上述代码创建的rho beta gamma 在反向传播的过程之中参数一直都维持着初始化时候的值

是我的用法不对吗