Class Flatten

code/chapter9.py:747–811

############### Flatten ##################
class Flatten(LayerBase):
    
    def __init__(self, keep_dim="first", optimizer=None):
        """
        将多维输入展开

        参数说明：
        keep_dim：展开形状，str (default : 'first')
                对于输入 X，keep_dim可选 'first'->将 X 重构为(X.shape[0], -1)，
                'last'->将 X 重构为(-1, X.shape[0])，'none'->将 X 重构为(1,-1)
        optimizer：优化方法
        """
        super().__init__(optimizer)

        self.keep_dim = keep_dim
        self._init_params()

    def _init_params(self):
        self.X = []
        self.gradients = {}
        self.params = {}
        self.derived_variables = {"in_dims": []}

    def forward(self, X, retain_derived=True):
        """
        前向传播

        参数说明：
        X：输入数组
        retain_derived：是否保留中间变量，以便反向传播时再次使用，bool型
        """
        if retain_derived:
            self.derived_variables["in_dims"].append(X.shape)
        if self.keep_dim == "none":
            return X.flatten().reshape(1, -1)
        rs = (X.shape[0], -1) if self.keep_dim == "first" else (-1, X.shape[-1])
        return X.reshape(*rs)

    def backward(self, dLdy, retain_grads=True):
        """
        反向传播

        参数说明：
        dLdy：关于损失的梯度
        retain_grads：是否计算中间变量的参数梯度，bool型

        输出说明：
        dX：将对输入的梯度进行重构为原始输入的形状
        """
        if not isinstance(dLdy, list):
            dLdy = [dLdy]
        in_dims = self.derived_variables["in_dims"]
        dX = [dy.reshape(*dims) for dy, dims in zip(dLdy, in_dims)]
        return dX[0] if len(dLdy) == 1 else dX

    @property
    def hyperparams(self):
        return {