如何实现keras中提出的tree-lstm，以及树结构是如何构建的？

import numpy as np
from keras.models import Model
from keras.layers import Input, Embedding, Dense
from keras.layers.recurrent import LSTM
from keras.layers.wrappers import TimeDistributed
from keras.layers.merge import concatenate

input_dim = 100 # 输入节点的维度
output_dim = 150 # 子节点的维度

node_input = Input(shape=(input_dim,)) # 输入节点
child_input = Input(shape=(input_dim,)) # 子节点

node_transform = Dense(output_dim)(node_input)

child_transform = Dense(output_dim)(child_input)

lstm = LSTM(output_dim)

node_lstm = lstm(node_transform)
child_lstm = lstm(child_transform)

merge = concatenate([node_lstm, child_lstm], axis=-1)

output = TimeDistributed(Dense(output_dim, activation='softmax'))(merge)

model = Model(inputs=[node_input, child_input], outputs=output)
model.compile(optimizer='adam', loss='categorical_crossentropy')

class TreeNode:
    def __init__(self, features, children):
        self.features = features # 节点特征
        self.children = children # 子节点列表

def build_tree(node):
    # 基准情况：如果节点没有子节点，则返回节点的特征
    if not node.children:
        return node.features
    
    # 递归情况：构建子节点的树结构，并将子节点的特征作为输入
    child_inputs = [build_tree(child) for child in node.children]
    
    # 将节点的特征和子节点的特征作为输入传递给模型进行预测
    inputs = [np.array([node.features]), np.array(child_inputs)]
    predictions = model.predict(inputs)
    
    # 返回预测结果
    return predictions[0]