用Softmax回归进行批量梯度下降训练---不使用Scikit-Learn

import numpy as np
from sklearn import datasets

def softmax(x, paras):
    value = np.exp(np.dot(x, paras))
    sum = np.sum(value)
    perc = value/sum
    return perc

iris = datasets.load_iris()
X = iris.data[:,2:4]
y = iris.target
# 添加一列为1的特征
X_bias =np.ones([len(X),1])
X_train_test = np.hstack((X_bias, X))
num_classes = 3
y_train_test = np.eye(num_classes)[y]
# print(iris.target.shape) # (150,)
#print(iris.data.shape) # (150, 4)
train_percent = 0.6

shuffle_index = np.random.permutation(150)
X_train_test = X_train_test[shuffle_index]
y_train_test = y_train_test[shuffle_index]
X_train = X_train_test[:int(150*train_percent)]
y_train = y_train_test[:int(150*train_percent)]
X_test = X_train_test[int(150*train_percent):]
y_test = y_train_test[int(150*train_percent):]
y_test_1 = np.argmax(y_test, axis=1)

def train_test(): 
    # 初始化参数
    para = np.random.rand(3,3)

    epoch = 20000
    samples = len(X_train)
    y_mid = np.zeros([int(150*train_percent),3])
    daoshu = np.zeros([3,3])
    loss = 0
    for i in range(epoch):
        loss = 0
        daoshu = np.zeros([3,3])
    #     print('daoshu=', daoshu)
        for j in range(samples):
            yj = y_train[j]
    #         print('yj=', yj)
    #         print('X_train[j]=', X_train[j])
            pj = softmax(X_train[j], para)
    #         print('pj=', pj)
    #         print('np.log(pj).T=', np.log(pj).T)
            loss += -np.dot(yj, np.log(pj).T)
    #         print('loss in=', loss)
            #daoshu = np.multiply(np.subtract(pj, yj), X_train[j])
            for k in range(3):
    #             print('np.subtract(pj[k], yj[k]) in=', np.subtract(pj[k], yj[k]))
    #             print('np.multiply(np.subtract(pj[k], yj[k]), X_train[j])=', np.multiply(np.subtract(pj[k], yj[k]), X_train[j]))
                daoshu[:, k] += np.multiply(np.subtract(pj[k], yj[k]), X_train[j]).T
    #             print('daoshu in=', daoshu)
            #print("daoshu in=", daoshu)
        loss = loss/samples
    #     print('loss out=', loss)
        daoshu = daoshu/samples
    #     print('daoshu out=', daoshu)
        para = para-np.multiply(0.01, daoshu)
    #     print('para out=', para)
        if (i+1)%2000 == 0:
            print('\ni=', i+1)
            print('loss=', loss)
            print('para=', para)
            
            y_pred = np.dot(X_test, para)
            y_pred = np.argmax(np.dot(X_test, para), axis=1)
            print(np.sum((y_pred==y_test_1).astype(int))/(150*(1-train_percent)))
            
            
train_test()

输出：

i= 2000
loss= 0.5050712902631536
para= [[ 2.06602856 -0.20443347 -1.32824731]
 [-0.22687971  0.34568269  0.24767398]
 [-0.41704432  0.65611776  1.96303604]]
0.6833333333333333

i= 4000
loss= 0.4003024132794443
para= [[ 2.9541613  -0.23096979 -2.18984373]
 [-0.40146026  0.48742211  0.28051512]
 [-0.75027279  0.42649992  2.52588234]]
0.7833333333333333

i= 6000
loss= 0.3454108424592416
para= [[ 3.57064547 -0.18903025 -2.84826744]
 [-0.52227931  0.57949177  0.3092645 ]
 [-0.97816569  0.24490013  2.93537503]]
0.8333333333333334

i= 8000
loss= 0.3091782305321345
para= [[ 4.0591322  -0.12135313 -3.40443129]
 [-0.61543022  0.64303481  0.33887237]
 [-1.15760138  0.10203338  3.25767747]]
0.8666666666666667

i= 10000
loss= 0.2825937078139275
para= [[ 4.47125948 -0.04349193 -3.89441976]
 [-0.69203841  0.68870641  0.36980897]
 [-1.30858544 -0.01173358  3.52242849]]
0.8666666666666667

i= 12000
loss= 0.2618906640587086
para= [[ 4.83136685  0.03787938 -4.33589844]
 [-0.75764403  0.72245435  0.40166665]
 [-1.4403523  -0.10371202  3.7461738 ]]
0.9166666666666666

i= 14000
loss= 0.24514135659598718
para= [[ 5.15302211  0.11959889 -4.73927321]
 [-0.8153586   0.74787342  0.43396214]
 [-1.55796442 -0.17916641  3.9392403 ]]
0.9166666666666666

i= 16000
loss= 0.2312253725585263
para= [[ 5.44469222  0.20009712 -5.11144156]
 [-0.86709636  0.7672674   0.46630593]
 [-1.66454441 -0.24188349  4.10853737]]
0.9166666666666666

i= 18000
loss= 0.2194325974010466
para= [[ 5.71209939  0.27859461 -5.45734622]
 [-0.91412081  0.78218025  0.49841752]
 [-1.76218754 -0.2946214   4.25891841]]
0.9166666666666666

i= 20000
loss= 0.20928408820673536
para= [[ 5.95934225  0.35472721 -5.78072168]
 [-0.95731318  0.79368403  0.53010612]
 [-1.85238907 -0.3394213   4.39391984]]
0.9166666666666666