paddle实现,多维时序数据增强 ，mixup（利用beta分布制作连续随机数）

#数据增强
def data_augment(X, y, p=0.8, alpha=0.5, beta=0.5):
    """Regression SMOTE
    1.将数据x分为 fix_X和X
    2.对X进行重塑
        对随机的均匀分布小于0.8的索引idx_to_change对应数据部分进行重塑。
        X2为打乱顺序后的X
        beta值为 np.random.beta(alpha, beta, batch_size) / 2 + 0.5 ####beta值>0.5 确保重塑数据中大部分是原值

        X[idx_to_change] = beta值*X[idx_to_change] +(1-beta值)*X2[idx_to_change]

    3.合并fix_X和X成新的数据x
    4.对于输入y同理处理
    """
    fix_X, X = X[:, :, :, :2], X[:, :, :, 2:]#[32, 134, 144, 2] [32, 134, 144, 10]
    fix_y, y = y[:, :, :, :2], y[:, :, :, 2:]#[32, 134, 288, 2] [32, 134, 288, 10]
    # print('fix_X, X:',fix_X.shape, X.shape)
    batch_size = X.shape[0]#32
    random_values = paddle.rand([batch_size]) #返回符合均匀分布的，范围在[0, 1)的Tensor  shape:[32]
    idx_to_change = random_values < p #32个值中，小于0.8的为False,其余为True

    # ensure that first element to switch has probability > 0.5
    np_betas = np.random.beta(alpha, beta, batch_size) / 2 + 0.5
    random_betas = paddle.to_tensor(
        np_betas, dtype="float32").reshape([-1, 1, 1, 1]) # [32, 1, 1, 1]
    index_permute = paddle.randperm(batch_size)# returns a 1-D Tensor filled with random permutation values from 0 to n-1#用于打乱数据编号

    X[idx_to_change] = random_betas[idx_to_change] * X[idx_to_change]
    X[idx_to_change] += (
        1 - random_betas[idx_to_change]) * X[index_permute][idx_to_change]

    y[idx_to_change] = random_betas[idx_to_change] * y[idx_to_change]
    y[idx_to_change] += (
        1 - random_betas[idx_to_change]) * y[index_permute][idx_to_change]
    return paddle.concat([fix_X, X], -1), paddle.concat([fix_y, y], -1)