tensorflow 实现游乐场的网络

图1 下面的tf程序用来搭建图中的网络结构, 并用同样的样本分布训练它.

# 下面的tf程序用来搭建 playground 中的网络结构, 并用同样的样本分布训练它# http://playground.tensorflow.org/import tensorflow as tffrom sklearn.datasets import make_circlesfrom sklearn.model_selection import train_test_splitfrom matplotlib import pyplot as pltN_SAMPLES=1000N = N_SAMPLES*(1-0.2)TEST_SIZE_PERCENT=0.2INPUT_DIMENSION = 2LEARNING_RATE = 0.03def get_trainset_and_testset():    X, y = make_circles(n_samples=N_SAMPLES, noise=0.2, factor=0.2, random_state=1)    y=y.reshape(N_SAMPLES,1)    for i in range(len(y)):        if y[i][0] == 0:            y[i][0] = -1    X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=.2)    return X_train, X_test, y_train, y_testdef plot_samples(X, y):    x1_axis_of_positive_sample = []    x2_axis_of_positive_sample = []    x1_axis_of_negative_sample = []    x2_axis_of_negative_sample = []    for i in range(len(y)):        if y[i][0] == 1:            x1_axis_of_positive_sample.append(X[i][0])            x2_axis_of_positive_sample.append(X[i][1])        else:            x1_axis_of_negative_sample.append(X[i][0])            x2_axis_of_negative_sample.append(X[i][1])    fig = plt.figure()    ax = fig.add_subplot(111)    type1 = ax.scatter(x1_axis_of_positive_sample, x2_axis_of_positive_sample, c='blue')    type2 = ax.scatter(x1_axis_of_negative_sample, x2_axis_of_negative_sample, c='brown')    plt.show()def tf_boy(X_train, y_train):    x = tf.placeholder(tf.float32, shape=(N, INPUT_DIMENSION), name="x-input")    y = tf.placeholder(tf.float32, shape=(N, 1), name='y-output')    w_12 = tf.Variable(tf.random_normal([INPUT_DIMENSION, 4]))    w_23 = tf.Variable(tf.random_normal([4, 2]))    w_34 = tf.Variable(tf.random_normal([2, 1]))    bias_2 = tf.Variable(tf.zeros([4]))    bias_3 = tf.Variable(tf.zeros([2]))    out_2 = tf.tanh(tf.matmul(x, w_12)+bias_2)    out_3 = tf.tanh(tf.matmul(out_2, w_23)+bias_3)    y_pred = tf.tanh(tf.matmul(out_3, w_34))    loss = tf.losses.mean_squared_error(y_pred, y_train)    train = tf.train.GradientDescentOptimizer(learning_rate=LEARNING_RATE).minimize(loss)    with tf.Session() as sess:        init_op = tf.global_variables_initializer()        sess.run(init_op)        # 输出目前（未经训练）的参数取值        print("w_12:", sess.run(w_12))        print("bias_2:", sess.run(bias_2))        print("w_23:", sess.run(w_23))        print("bias_3:", sess.run(bias_3))        print("w_34:", sess.run(w_34))        print("\n")        # 训练模型        STEPS = 1000        loss_arr=[]        for i in range(STEPS):            sess.run(train, feed_dict={x: X_train, y: y_train})            # print("w_12:", sess.run(w_12))            total_loss = sess.run(loss, feed_dict={x: X_train, y: y_train})            loss_arr.append(total_loss)        # 输出训练后的参数取值。        print("\n")        print("w_12:", sess.run(w_12))        print("bias_2:", sess.run(bias_2))        print("w_23:", sess.run(w_23))        print("bias_3:", sess.run(bias_3))        print("w_34:", sess.run(w_34))        #损失函数值, 变化曲线        fig = plt.figure()        ax = fig.add_subplot(111)        x1_axis=range(STEPS)        x2_axis=loss_arr        ax.scatter(x1_axis, x2_axis, c='blue')        plt.show()X_train, X_test, y_train, y_test = get_trainset_and_testset()plot_samples(X_train, y_train)tf_boy(X_train, y_train)"""w_12: [[-1.02140009  0.1605026  -1.79282069  0.4733575 ] [ 0.39666429  0.18492372 -0.76060712  1.30654573]]bias_2: [ 0.  0.  0.  0.]w_23: [[ 0.67617249  0.37842372] [-0.13533577 -0.66545761] [-1.15543711 -2.49448705] [ 2.22956133 -0.32165667]]bias_3: [ 0.  0.]w_34: [[-0.3632547 ] [ 1.46542287]]w_12: [[-1.49568939  0.37617555 -1.63048255  0.28560442] [ 0.9723621   0.18761159 -0.50179791  1.75902331]]bias_2: [ 0.82246989  0.85552055 -0.75419569 -1.0005064 ]w_23: [[ 0.47892851  1.46363699] [-0.24798357 -1.31343782] [-0.90114665 -1.79665875] [ 2.36001635 -1.09818363]]bias_3: [-0.33126104 -1.38699543]w_34: [[-0.48449364] [ 2.55953455]]"""