tensorflowにおける乱数の固定について

Question

tensorflowで乱数を発生させているが，固定することが出来ない．

ハイパーパラメータである学習率を4通り考慮した．4回ループを回すことになるが，固定されているはずの乱数が各ループでずれてしまっている．
今後，グリッドサーチを行おうと考えているのでとても困っています．どのように解決すべきでしょうか．
質問・プログラミングについて初心者でりますがお願いします．

発生している問題・エラーメッセージ

0.01
[[-0.01278295]
 [ 0.00424424]
 [ 0.01260157]
 [ 0.00272857]
 [ 0.00025477]
 [-0.00312508]
 [-0.00225875]
 [-0.00445066]
 [ 0.00055921]
 [-0.00066416]
 [ 0.00816493]
 [ 0.00378237]]
0.05
[[-0.00315907]
 [-0.00032044]
 [ 0.00632677]
 [ 0.01084339]
 [ 0.0176496 ]
 [-0.00058672]
 [ 0.00848717]
 [ 0.00399379]
 [ 0.00903138]
 [-0.00502061]
 [-0.00796483]
 [-0.00799216]]
0.1
[[ 0.00535443]
 [ 0.00195984]
 [-0.00651257]
 [-0.00675841]
 [-0.01028436]
 [-0.01944688]
 [-0.00674361]
 [ 0.00747328]
 [-0.00119065]
 [-0.00600063]
 [-0.00499562]
 [ 0.01717699]]
0.25
[[ 0.01153731]
 [-0.0178353 ]
 [ 0.00059123]
 [ 0.01206309]
 [ 0.00179897]
 [-0.00935876]
 [ 0.00936141]
 [-0.00026272]
 [ 0.01812616]
 [-0.00089715]
 [ 0.00351766]
 [-0.00794965]]
[<matplotlib.lines.Line2D at 0x1023ffd0>]

該当のソースコード

import numpy as np
import tensorflow as tf
%matplotlib inline
import matplotlib.pyplot as plt
with tf.Graph().as_default():
    randomseed=1234
    #np.random.seed(randomseed)
    tf.set_random_seed(randomseed)
    n_in = 20
    n_hidden = 12
    #n_hidden = [1,2,3,4,5,6,7,8,9,10]
    n_out = 1
    #学習回数
    epochs=10
    #epochs=[]
    #学習率
    #LEARNING_RATE=0.01
    LEARNING_RATE=[0.01,0.05,0.10,0.25]
    for learn in LEARNING_RATE:
        randomseed=1234
        #np.random.seed(randomseed)
        tf.set_random_seed(randomseed)
    
    #学習
    #学習モデルの設定
    #tensorflowの定義
        x = tf.placeholder(dtype= tf.float32, shape=[None,n_in], name='x')
        t = tf.placeholder(dtype= tf.float32, shape=[None,n_out], name='t')
    
    #入力層から中間層
        w21 = tf.Variable(tf.truncated_normal([n_in, n_hidden], stddev=0.01))
        b21 = tf.Variable(tf.zeros([n_hidden]))
        h21 = tf.nn.sigmoid(tf.matmul(x, w21) + b21)
    #中間層から出力層
        w32 = tf.Variable(tf.truncated_normal([n_hidden, n_out], stddev=0.01))
        b32 = tf.Variable(tf.zeros([n_out]))
        y = tf.nn.sigmoid(tf.matmul(h21, w32) + b32)
    #2乗誤差
        loss = tf.reduce_sum(tf.square(y - t))
    #最急降下法
        train_step = tf.train.GradientDescentOptimizer(learn).minimize(loss)
    #ここまでモデルの定義
    #モジュールのインストール

        with tf.Session() as sess:
            randomseed=1234
            #np.random.seed(randomseed)
            tf.set_random_seed(randomseed)
            sess.run(tf.global_variables_initializer())
            randomseed=1234
            #np.random.seed(randomseed)
            tf.set_random_seed(randomseed)
            
            loss_cal=[]
            print(learn)
            print(sess.run(w32))
　　　　　　　for _ in range(epochs):
            　y_out, _ = sess.run([loss, train_step], feed_dict = {
            　x: inputdata[:],
            　t: outputdata[:]
            　})
    

試したこと

どこに乱数固定のプログラムを書いたとしても固定されない．

Answer 1

LEARNING_RATEの繰り返しをw21，w32など乱数の定義より下にすることで乱数の固定は出来ました．
しかし，w21やw32のn_hiddenを何パターンか変える場合同じ手は使えません．

with tf.Graph().as_default():
randomseed=1234
np.random.seed(randomseed)
tf.set_random_seed(randomseed)
n_in = 20
n_hidden = 12
#n_hidden = [1,2,3,4,5,6,7,8,9,10]
n_out = 1
epochs=10
x = tf.placeholder(dtype= tf.float32, shape=[None,n_in], name='x')
t = tf.placeholder(dtype= tf.float32, shape=[None,n_out], name='t')

w21 = tf.Variable(tf.truncated_normal([n_in, n_hidden], stddev=0.01))
b21 = tf.Variable(tf.zeros([n_hidden]))
h21 = tf.nn.sigmoid(tf.matmul(x, w21) + b21)
w32 = tf.Variable(tf.truncated_normal([n_hidden, n_out], stddev=0.01))
b32 = tf.Variable(tf.zeros([n_out]))
y = tf.nn.sigmoid(tf.matmul(h21, w32) + b32)
loss = tf.reduce_sum(tf.square(y - t))

LEARNING_RATE=[0.01,0.05,0.10,0.25]
for learn in LEARNING_RATE:
    train_step = tf.train.GradientDescentOptimizer(learn).minimize(loss)
    with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())
        loss_cal=[]
        print(learn)
        print(sess.run(w32))
        
        for _ in range(epochs):
            y_out, _ = sess.run([loss, train_step], feed_dict = {
            x: inputdata[:],
            t: outputdata[:]
            })
            loss_cal.append(y_out)

plt.plot(loss_cal)

0.01
[[-0.01278295]
[ 0.00424424]
[ 0.01260157]
[ 0.00272857]
[ 0.00025477]
[-0.00312508]
[-0.00225875]
[-0.00445066]
[ 0.00055921]
[-0.00066416]
[ 0.00816493]
[ 0.00378237]]
0.05
[[-0.01278295]
[ 0.00424424]
[ 0.01260157]
[ 0.00272857]
[ 0.00025477]
[-0.00312508]
[-0.00225875]
[-0.00445066]
[ 0.00055921]
[-0.00066416]
[ 0.00816493]
[ 0.00378237]]
0.1
[[-0.01278295]
[ 0.00424424]
[ 0.01260157]
[ 0.00272857]
[ 0.00025477]
[-0.00312508]
[-0.00225875]
[-0.00445066]
[ 0.00055921]
[-0.00066416]
[ 0.00816493]
[ 0.00378237]]
0.25
[[-0.01278295]
[ 0.00424424]
[ 0.01260157]
[ 0.00272857]
[ 0.00025477]
[-0.00312508]
[-0.00225875]
[-0.00445066]
[ 0.00055921]
[-0.00066416]
[ 0.00816493]
[ 0.00378237]]