numpy, list の append の速度の違い

やること

numpy と python list の append が結構速度が違うと見たので、実際に違いを計測してみる。

コード

import numpy as np
import time
import matplotlib.pyplot as plt
import pandas as pd

data_nums = np.logspace(1, 5, 5)
# [1.e+01 1.e+02 1.e+03 1.e+04 1.e+05]
data_nums = data_nums.astype("int64")

df = pd.DataFrame(
    index=range(len(data_nums)), columns=["data_num", "python_list", "numpy_ndarray"]
)
df.loc[:, "data_num"] = data_nums

for data_num in data_nums:

    df_index = df.index[df["data_num"] == data_num].to_list()
    print(f"-" * 30)
    print(f"data_num : {data_num}")
    print(f"-" * 30)

    # =========================================
    # using python list append
    # =========================================
    start_time_python_list = time.perf_counter()

    data_python_list = np.empty(0)
    data_python_list_ = data_python_list.tolist()  #  numpy ndarray -> python list
    for i in range(data_num):  # append data
        data_python_list_.append(i)
    data_python_list_np = np.asarray(data_python_list_)
    # python list -> numpy ndarray
    # print(data_python_list_np)

    end_time_python_list = time.perf_counter()
    elapsed_time_python_list = end_time_python_list - start_time_python_list
    print(f"python_list : {elapsed_time_python_list} [sec]")
    df.at[df.index[df_index[0]], "python_list"] = elapsed_time_python_list

    # =========================================
    # using numpy ndarray append
    # =========================================
    start_time_numpy = time.perf_counter()

    data_numpy_append_ = np.empty(0)
    for i in range(data_num):  # append data
        data_numpy_append_ = np.append(data_numpy_append_, i)
    # print(data_numpy_append_)

    end_time_numpy = time.perf_counter()
    elapsed_time_numpy = end_time_numpy - start_time_numpy
    print(f"numpy_ndarray : {elapsed_time_numpy} [sec]")
    df.at[df.index[df_index[0]], "numpy_ndarray"] = elapsed_time_numpy

    print(f"")

print(df)
a = df.to_numpy()

# グラフ描画用
fig = plt.figure()
ax = fig.add_subplot(1, 1, 1)
ax.set_xscale("log")
ax.set_xlabel("data num")
ax.set_ylabel("elapsed time[s]")
ax.plot(a[:, 0], a[:, 1], label="python_list")
ax.scatter(a[:, 0], a[:, 1])
ax.plot(a[:, 0], a[:, 2], label="numpy_ndarray")
ax.scatter(a[:, 0], a[:, 2])
ax.legend()
plt.show()

結果

配列の要素数（data_num）	python_list [sec]	numpy_ndarray [sec]
10	0.000019	0.000081
100	0.000221	0.000634
1000	0.00026	0.007898
10000	0.001955	0.071762
100000	0.018196	2.05108

append は python list の方が速い。

numpy ndarray は固定長で宣言されており、python list は可変長配列であることが理由らしい。