点到点和图到图带代码

点到点的代码理解

点击查看代码

def load_train_data_for_rnn(cfg, x, y, aux, scaler):
    # x = {nt, nf, ngrid} = {3287,9,1399}
    # y = {nt, ngrid} = {3287,1399}
    # aux = {nf , nt} = {1 , 1399}
    # scaler = {2, 45, 90, 1 }
    nt, nf, ngrid = x.shape
    # print('**************************, y.shape is ' + str(y.shape))
    # print('**************************, aux.shape is ' + str(aux.shape))
    # print('**************************, x.shape is ' + str(x.shape))
    # print('**************************, scaler.shape is ' + str(scaler.shape))
    mean, std = np.array(scaler[0]), np.array(scaler[1])
    # print('**************************, mean.shape is ' + str(mean.shape))
    #mean = mean.reshape(mean.shape[0],mean.shape[1]*mean.shape[2])
    #std = std.reshape(std.shape[0],std.shape[1]*std.shape[2])
    idx_time = np.random.randint(0, nt-cfg['seq_len']-cfg["forcast_time"], 1)[0]#在这个代码中，由于最后使用了索引 [0]，所以实际上只生成一个随机整数。 1 作为参数传递给了 np.random.randint()，所以实际上只生成了一个随机整数
    # print('**************************, idx_time is ' + str(idx_time))
    idx_grid = np.random.randint(0, ngrid, cfg['batch_size'])#会生成 cfg['batch_size'] 个位于 0（包含）和 ngrid（不包含）之间的随机整数,数组形式，大小是64，它生成了 cfg['batch_size'] 个随机整数
    # print('**************************, idx_grid is ' + str(idx_grid))
    # print('**************************, idx_grid.shape is ' + str(idx_grid.shape))======idx_grid，64個（0到1399）的數組
    x = np.transpose(x, (2,0,1))#x is {ngrid,nt,nf} = { 1399, 3287 , 9 }
    y = np.transpose(y, (1,0))# y is {ngrid,nt} = {1399 , 3287}
    aux = np.transpose(aux, (1,0)) # aux is {nt, nf}={1399,1}
    # print('before select form x **************************, x is ' + str(x.shape))#========x{1399,3287,9}
    x = x[idx_grid, idx_time:idx_time+cfg['seq_len']]
    # print('after selcect from x **************************, x is ' + str(x.shape))=========x {64,365,9}
    # print('before select form y **************************, y is ' + str(y.shape))# y is (1399, 3287)
    y = y[idx_grid, idx_time+cfg['seq_len']+cfg["forcast_time"]] ##
    # print('after selcect from y **************************, y is ' + str(y.shape))# y is (64,)
    aux = aux[idx_grid]
    y[np.isinf(y)]=np.nan#将这些选中的无穷大值替换为 NaN
    mask = y == y#创建一个布尔类型的掩码（mask），判断数组 y 中的元素是否为有效值
    x = x[mask]
    y = y[mask]
    aux = aux[mask]
    x[np.isinf(x)]=np.nan
    x = np.nan_to_num(x)#将数组 x 中的 NaN 值替换为 0。
    return x, y, aux, mean, std