为什么numpy.fip()在不改变其精度和形状的情况下减少numpy数组的内存大小

我编写了一个函数,将序列数据转换为稀疏矩阵。我惊讶地发现 img = np.flip(img, axis = 0) 将矩阵存储器的大小从1928减小到128。

此外,无论我如何增加时间窗口的长度(稀疏矩阵的宽度),最终翻转矩阵的内存大小始终为128。

我以为 np.flip() 只创建了原始矩阵的视图,这意味着内存大小应该更改,所以我不明白为什么会发生这种情况。

函数如下所示

def gen_img(df, index, interval, height):
    df = df[index-interval+1: index+1, :5]
    max_price = df[:, 1].max()
    min_price = df[:, 2].min()
    if max_price == min_price:
        return df, np.zeros((height, interval * 3), dtype="float32")
    else:
        df[:, :4] = np.around((height * 4 / 5) * (df[:, :4] - min_price) / (max_price - min_price) + height / 5)
        df[:, 4] = np.around(df[:, 4].astype("float32") * height / 5 / df[:, 4].max())
        df = df.astype(int)
        img = np.zeros((height, interval * 3), dtype="float32")
        print("check zero matrix size: {}".format(sys.getsizeof(img)))
        for i in range(interval):
            img[df[i, 0], i*3] = 1.0
            print("check intermed1 matrix size: {}".format(sys.getsizeof(img)))
            img[df[i, 3], i*3 + 2] = 1.0
            img[df[i, 2] : df[i, 1] + 1, i*3 + 1] = 1.0
            img[0 : df[i, 4] + 1, i*3 + 1] = 1.0
        print("check intermed2 matrix size: {}".format(sys.getsizeof(img)))
        img = np.flip(img, axis = 0)
        print("check generated matrix size: {}".format(sys.getsizeof(img)))

    return df, img

这是调试打印结果

check zero matrix size: 1928, dtype: float32, shape: (30, 15)
check intermed1 matrix size: 1928, dtype: float32, shape: (30, 15)
check intermed1 matrix size: 1928, dtype: float32, shape: (30, 15)
check intermed1 matrix size: 1928, dtype: float32, shape: (30, 15)
check intermed1 matrix size: 1928, dtype: float32, shape: (30, 15)
check intermed1 matrix size: 1928, dtype: float32, shape: (30, 15)
check intermed2 matrix size: 1928, dtype: float32, shape: (30, 15)
check generated matrix size: 128, dtype: float32, shape: (30, 15)