STGCN复现第三弹：解读math_graph.py

import numpy as np
import pandas as pd
from scipy.sparse.linalg import eigs

scaled_laplacian函数：

def scaled_laplacian(W):
    '''
    Normalized graph Laplacian function. 标准化拉普拉斯矩阵
    :param W: np.ndarray, [n_route, n_route], weighted adjacency matrix of G.
    :return: np.matrix, [n_route, n_route].
    '''
    # d ->  diagonal degree matrix 度矩阵
    n, d = np.shape(W)[0], np.sum(W, axis=1)      #1.取出矩阵W的维度 2.将矩阵W的列向量相加
    ’‘’
    举例：
    c = np.array([[0, 2, 1], [3, 5, 6], [0, 1, 1]])
	print c.sum()
	print c.sum(axis=0)
	print c.sum(axis=1)
	结果分别是：19, [3 8 8], [ 3 14  2]
    ‘’‘
    # L -> graph Laplacian
    L = -W
    L[np.diag_indices_from(L)] = d  #返回索引以访问n维数组L的主对角线
    for i in range(n):
        for j in range(n):
            if (d[i] > 0) and (d[j] > 0):
                L[i, j] = L[i, j] / np.sqrt(d[i] * d[j])
    # lambda_max \approx（大约） 2.0, the largest eigenvalues（特征值） of L.
    lambda_max = eigs(L, k=1, which='LR')[0][0].real  # 1. k：int, 可选参数所需的特征值和特征向量的数量。 2.‘LR’:largest real part 3. x.real取得xd的实部
     #eigs（）全称为scipy.sparse.linalg.eigs（）用于求平方矩阵A的k个特征值和特征向量。
    return np.mat(2 * L / lambda_max - np.identity(n)) 	#np.identity(n)生成一个n行n列的单位矩阵

np.identity()函数详解
scipy.sparse.linalg.eigs（）函数详解~

cheb_poly_approx函数：

def cheb_poly_approx(L, Ks, n):
    '''
    Chebyshev(切比雪夫) polynomials（多项式） approximation（逼近、近似） function.
    :param L: np.matrix, [n_route, n_route], graph Laplacian（拉普拉斯算子）.
    :param Ks: int, kernel size of spatial convolution（空间卷积核大小）.
    :param n: int, number of routes / size of graph.（图的大小）
    :return: np.ndarray, [n_route, Ks*n_route]. 返回一个np.ndarray的多维数组
    '''
    L0, L1 = np.mat(np.identity(n)), np.mat(np.copy(L))  	#np.copy():返回给定对象的数组副本

    if Ks > 1:
        L_list = [np.copy(L0), np.copy(L1)]
        for i in range(Ks - 2):
            Ln = np.mat(2 * L * L1 - L0)
            L_list.append(np.copy(Ln))
            L0, L1 = np.matrix(np.copy(L1)), np.matrix(np.copy(Ln))
        # L_lsit [Ks, n*n], Lk [n, Ks*n]
        return np.concatenate(L_list, axis=-1)
    elif Ks == 1:
        return np.asarray(L0) 	#将输入转换为数组
    else:
        raise ValueError(f'ERROR: the size of spatial kernel must be greater than 1, but received "{Ks}".')

np.copy()函数详解
np.concatenate()函数详解
numpy.asarray（）函数详解

first_approx()函数：

def first_approx(W, n):
    '''
    1st-order approximation function.
    :param W: np.ndarray, [n_route, n_route], weighted adjacency matrix of G.
    :param n: int, number of routes / size of graph.
    :return: np.ndarray, [n_route, n_route].
    '''
    A = W + np.identity(n)
    d = np.sum(A, axis=1)
    sinvD = np.sqrt(np.mat(np.diag(d)).I)
    # refer to Eq.5
    return np.mat(np.identity(n) + sinvD * A * sinvD)

weight_matrix（）函数：

def weight_matrix(file_path, sigma2=0.1, epsilon=0.5, scaling=True):
    '''
    Load weight matrix function.
    :param file_path: str, the path of saved weight matrix file.
    :param sigma2: float, scalar of matrix W.
    :param epsilon: float, thresholds to control the sparsity of matrix W.
    :param scaling: bool, whether applies numerical scaling on W.
    :return: np.ndarray, [n_route, n_route].
    '''
    try:
        W = pd.read_csv(file_path, header=None).values
    except FileNotFoundError:
        print(f'ERROR: input file was not found in {file_path}.')

    # check whether W is a 0/1 matrix.
    if set(np.unique(W)) == {0, 1}:
        print('The input graph is a 0/1 matrix; set "scaling" to False.')
        scaling = False

    if scaling:
        n = W.shape[0]
        W = W / 10000.
        W2, W_mask = W * W, np.ones([n, n]) - np.identity(n)
        # refer to Eq.10
        return np.exp(-W2 / sigma2) * (np.exp(-W2 / sigma2) >= epsilon) * W_mask
    else:
        return W