機器學習基石作业1 Q15-17 感知器学习算法 (PLA) 的Python实现

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
%matplotlib inline

names = ["x1", "x2", "x3", "x4", "y"]
df = pd.read_csv("~/Documents/PLA.csv", names=names)
df.insert(0, "x0", 1)

timels=[]
for i in range(100):
    w = pd.Series([0, 0, 0, 0, 0], index=["x0", "x1", "x2", "x3", "x4"]); times = 1; # wl = []
    # wl.append(w)
    mis_p = df[df["y"]==+1][np.dot(df[df["y"]==+1].iloc[:,0:df.shape[1]-1], w) <= 0]
    mis_n = df[df["y"]==-1][np.dot(df[df["y"]==-1].iloc[:,0:df.shape[1]-1], w) > 0]
    mis = pd.concat([mis_p, mis_n], axis=0)
    length = mis.shape[0]
    while length > 0:
        mis_point = mis.iloc[np.random.randint(length),:]
        w = w + mis_point["y"] * mis_point.iloc[0: mis_point.shape[0]-1]
        mis_p = df[df["y"]==+1][np.dot(df[df["y"]==+1].iloc[:,0:df.shape[1]-1], w) <= 0]
        mis_n = df[df["y"]==-1][np.dot(df[df["y"]==-1].iloc[:,0:df.shape[1]-1], w) > 0]
        mis = pd.concat([mis_p, mis_n], axis=0)
        length = mis.shape[0]
        times += 1; # wl.append(w)
        if times > 50000 :
            print("over 50000 times loops")
            break
    timels.append(times)

ts = pd.Series(timels)
ts.plot()
print(ts.mean())

平均 42.16 次

另外发现一个有意思的点:

w = pd.Series([0, 0, 0, 0, 0])
np.dot(df[df["y"]==+1].iloc[:,0:df.shape[1]-1], w)  # 可行
w.dot(df[df["y"]==+1].iloc[:,0:df.shape[1]-1]) # 报错