pytorch 循环生成网络结构

由于网络通常非常深，我们一层一层写会耗费很大的精力。而且修改起来也比较麻烦，不方便维护。网络通常由一些相似的块组成，我们写出一个块，循环生成整个网络。

nn.Sequential

def __init__(self, num_layers=1, activation=torch.nn.ReLU()):
    super(Classifier3, self).__init__()
    self.num_layers = num_layers
    self.dense1 = torch.nn.Sequential(
        torch.nn.utils.spectral_norm(torch.nn.Linear(512, 256)),
        activation
    )
    self.dense = self._make_layers(num_layers,activation)
    self.dense2 = torch.nn.Sequential(
        torch.nn.utils.spectral_norm(torch.nn.Linear(256, 128)),
        activation
    )
    self.dense3 = torch.nn.Sequential(
        torch.nn.utils.spectral_norm(torch.nn.Linear(128,64)),
        activation,
        torch.nn.Linear(64, 2),
        torch.nn.LogSoftmax(dim=1)
    )

def _make_layers(self, num_layers=1, activation=torch.nn.ReLU()):
    dense = []
    for i in range(num_layers):
        dense.append(torch.nn.Sequential(
            torch.nn.utils.spectral_norm(torch.nn.Linear(256, 256)),
            activation
        ))
    return torch.nn.Sequential(*dense)

nn.ModuleList

def __init__(self, num_layers=1, activation=torch.nn.ReLU()):
    super(Classifier3, self).__init__()
    self.num_layers = num_layers
    self.dense1 = torch.nn.Sequential(
        torch.nn.utils.spectral_norm(torch.nn.Linear(512, 256)),
        activation
    )
    dense = []
    for i in range(num_layers):
        dense.append(torch.nn.Sequential(
            torch.nn.utils.spectral_norm(torch.nn.Linear(256, 256)),
            activation
        ))
    self.dense_block = torch.nn.ModuleList(dense)
    self.dense2 = torch.nn.Sequential(
        torch.nn.utils.spectral_norm(torch.nn.Linear(256, 128)),
        activation
    )
    self.dense3 = torch.nn.Sequential(
        torch.nn.utils.spectral_norm(torch.nn.Linear(128,64)),
        activation,
        torch.nn.Linear(64, 2),
        torch.nn.LogSoftmax(dim=1)
    )