Created by: XDang13

class SpertralNormLinear(fluid.dygraph.Layer):
    def __init__(self, input_dim, dim, bias_attr=False):
        super(SpertralNormLinear, self).__init__()
        self.linear = Linear(input_dim, dim, bias_attr=bias_attr)
        self.sp_block = fluid.dygraph.nn.SpectralNorm(self.linear.weight.shape, dim=0, power_iters=2)
    def forward(self, input):
        self.spectral_norm()
        output = self.linear(input)
        return output
    def spectral_norm(self):
        weight = self.linear.weight
        norm_weight = self.sp_block(weight)
        self.linear.weight.set_value(norm_weight)

还没在GAN中测试，理论上应该是可以的。