我什么时候应该使用matrix.astype以及何时使用matrix.view？

Question

我有一个int32整数矩阵（50k x 50k左右），需要将其转换为float32。我可以做到

# Preparation for the example
import numpy as np
n = 50_000
matrix = np.random.randint(0, 10, (n, n), dtype='int32')

# Way 1:
matrix = matrix.astype(np.float32, copy=False)

# Way 2:
matrix = matrix.view(np.float32)

我什么时候应该使用哪个？与“实际”的numpy数组相比，在以后使用视图时是否存在速度上的劣势？

我尝试过的

执行时间分析（创建，以后不再访问）

import numpy as np
import timeit


def create_boxplot(duration_list, showfliers=False):
    import seaborn as sns
    import matplotlib.pyplot as plt
    import operator

    plt.figure(num=None, figsize=(8, 4), dpi=300, facecolor="w", edgecolor="k")
    sns.set(style="whitegrid")
    sorted_keys, sorted_vals = zip(
        *sorted(duration_list.items(), key=operator.itemgetter(1))
    )
    flierprops = dict(markerfacecolor="0.75", markersize=1, linestyle="none")
    ax = sns.boxplot(
        data=sorted_vals,
        width=0.3,
        orient="h",
        flierprops=flierprops,
        showfliers=showfliers,
    )
    ax.set(xlabel="Time in ms", ylabel="")
    plt.yticks(plt.yticks()[0], sorted_keys)
    plt.tight_layout()
    plt.savefig("output.png")


n = 5_000
matrix = np.random.randint(0, 2, (n, n), dtype='int32')
print(matrix.dtype)

matrix = matrix.view(np.float32)
print(matrix.dtype)

timeit_d = {}
timeit_d["repeat"] = 500
timeit_d["number"] = 3
timeit_d["setup"] = "import numpy as np; n=5_000; matrix = np.random.randint(0, 2, (n, n), dtype='int32')"

duration_list = {}

# Way 1
durations = timeit.repeat(
    "matrix2 = matrix.view(np.float32)",
    setup=timeit_d["setup"],
    repeat=timeit_d["repeat"],
    number=timeit_d["number"],
)
duration_list["view"] = durations
print("Done views")

# Way 2
durations = timeit.repeat(
    "matrix2 = matrix.astype(np.float32)",
    setup=timeit_d["setup"],
    repeat=timeit_d["repeat"],
    number=timeit_d["number"],
)
duration_list["astype"] = durations
print("Done astype")

# Visualize
create_boxplot(duration_list)

很明显，视图比astype快。

内存分析

$ valgrind --tool=massif python3 foobar.py
$ massif-visualizer massif.out.view

清楚地表明，view选项使用的内存更少。