TensorFlowで平均と分散の最尤推定

Tensorflowが機械学習ライブラリという誤解がまかり通っているので，理解を深めるために平均と分散を求めてみた．

• 平均50標準偏差10の正規乱数を100個生成．
• 学習係数をいつくか変更

平均だけならTensorFlowで数列の平均を求めてみた
を参照．

Gradient decent

import matplotlib.pylab as plt
%matplotlib inline
import numpy as np
import tensorflow as tf

x_train = np.random.randn(100) * 10 + 50

n_itr = 10000

m = tf.Variable([30.0], tf.float32)
s = tf.Variable([3.0], tf.float32)
x = tf.placeholder(tf.float32)
N = tf.count_nonzero(x, dtype=tf.float32)

loss = N / 2 * tf.log(2 * np.pi * s**2) + 1.0 / (2 * s**2) * tf.reduce_sum(tf.square(x - m))

for lr in [0.1, 0.01, 0.001]:

    optimizer = tf.train.GradientDescentOptimizer(lr)
    train = optimizer.minimize(loss)

    init = tf.global_variables_initializer()
    sess = tf.Session()
    sess.run(init)

    est_m = []
    est_s = []
    for i in range(n_itr):
        sess.run(train, {x:x_train})
        est_m.append(sess.run([m]))
        est_s.append(sess.run([s]))

    est_m = np.array(est_m)
    est_s = np.array(est_s)
    plt.plot(est_s.reshape(n_itr)[::100], est_m.reshape(n_itr)[::100], marker=".", label="lr={}".format(lr))

plt.title("batch gradient decent")
plt.xlabel("std")
plt.ylabel("mean")
plt.legend()
plt.show();

RMS Prop

n_itr = 1000

for lr in [1, 0.5, 0.1]:

    optimizer = tf.train.RMSPropOptimizer(lr)
    train = optimizer.minimize(loss)

    init = tf.global_variables_initializer()
    sess = tf.Session()
    sess.run(init)

    est_m = []
    est_s = []
    for i in range(n_itr):
        sess.run(train, {x:x_train})
        est_m.append(sess.run([m]))
        est_s.append(sess.run([s]))

    est_m = np.array(est_m)
    est_s = np.array(est_s)
    plt.plot(est_s.reshape(n_itr)[::10], est_m.reshape(n_itr)[::10], marker=".", label="lr={}".format(lr))

plt.title("batch RMS Prop")
plt.xlabel("std")
plt.ylabel("mean")
plt.legend()
plt.show();

Adam

n_itr = 1000

for lr in [5, 1, 0.1, 0.01]:

    optimizer = tf.train.AdamOptimizer(lr)
    train = optimizer.minimize(loss)

    init = tf.global_variables_initializer()
    sess = tf.Session()
    sess.run(init)

    est_m = []
    est_s = []
    for i in range(n_itr):
        sess.run(train, {x:x_train})
        est_m.append(sess.run([m]))
        est_s.append(sess.run([s]))

    est_m = np.array(est_m)
    est_s = np.array(est_s)
    plt.plot(est_s.reshape(n_itr)[::10], est_m.reshape(n_itr)[::10], marker=".", label="lr={}".format(lr))

plt.title("batch Adam")
plt.xlabel("std")
plt.ylabel("mean")
plt.legend()
plt.show();

AdaGrad

n_itr = 1000

for lr in [20, 10, 5, 1]:

    optimizer = tf.train.AdagradOptimizer(lr)
    train = optimizer.minimize(loss)

    init = tf.global_variables_initializer()
    sess = tf.Session()
    sess.run(init)

    est_m = []
    est_s = []
    for i in range(n_itr):
        sess.run(train, {x:x_train})
        est_m.append(sess.run([m]))
        est_s.append(sess.run([s]))

    est_m = np.array(est_m)
    est_s = np.array(est_s)
    plt.plot(est_s.reshape(n_itr)[::10], est_m.reshape(n_itr)[::10], marker=".", label="lr={}".format(lr))

plt.title("batch AdaGrad")
plt.xlabel("std")
plt.ylabel("mean")
plt.legend()
plt.show();

AdaDelta

n_itr = 1000

for lr in [5000, 1000, 100]:

    optimizer = tf.train.AdadeltaOptimizer(lr)
    train = optimizer.minimize(loss)

    init = tf.global_variables_initializer()
    sess = tf.Session()
    sess.run(init)

    est_m = []
    est_s = []
    for i in range(n_itr):
        sess.run(train, {x:x_train})
        est_m.append(sess.run([m]))
        est_s.append(sess.run([s]))

    est_m = np.array(est_m)
    est_s = np.array(est_s)
    plt.plot(est_s.reshape(n_itr)[::10], est_m.reshape(n_itr)[::10], marker=".", label="lr={}".format(lr))

plt.title("batch AdaDelta")
plt.xlabel("std")
plt.ylabel("mean")
plt.legend()
plt.show();