perceptron de tensorflow donne une sortie unexplaineble

Question

Je suis nouveau à TF: Je pris le code de perceptron de ce tutoriel sur MNIST (en fait, ce ne est pas nécessaire de suivre ce lien): https://github.com/aymericdamien/TensorFlow-Examples/blob/master/examples/3_NeuralNetworks/multilayer_perceptron.py

Je voulais refaire ces perceptron à un perceptron avec 1 couche et fonction d'activation linéaire, pour en faire la forme la plus simple de: output = w2 (w1 * x + b1) + b2. Mais voici ce que je reçois:

données:

X_train: Array ([[10], [10.], [11], [6], [8.], [9], [22], [14], [6], [8.], [11], [9], [13 .], [7.], [ 13.], [7.], [13.], [11.]])

y_train: Array ([[44,5825], [53,99], [52,4475], [37,6] , [38,6125], [39,5875], [43,07], [74,8575], [34,185], [38,61], [34,8175], [36,61], [34,0675], [37,67] , [49.725], [79.4775], [50,41], [51,26]])

X_test: Array ([[6], [14], [14], [12.], [13], [ 13.]])

y_test: array ([[55,75], [33,035], [38,3275], [39,2825], [50,7325], [45.2575]])

Paramètres:

learning_rate = 1 
training_epochs = 1 
display_step = 1 #maintaining variable 
x = tf.placeholder("float", [None, 1]) 
y = tf.placeholder("float", [None, 1]) 

modèle Perceptron:

def multilayer_perceptron(x, weights, biases, output_0): 
    layer_1 = tf.add(tf.matmul(x, weights['h1']), biases['b1']) 
    out_layer = tf.add(tf.matmul(layer_1, weights['out']), biases['out']) 
    output_o = out_layer #This variable is just needed to print result in session 
    return out_layer 

output_0 = tf.Variable(tf.random_normal([1, n_classes])) 
weights = { 
    'h1': tf.Variable(tf.random_normal([n_input, n_hidden_1])), 
    'out': tf.Variable(tf.random_normal([n_hidden_2, n_classes]))} 
biases = { 
    'b1': tf.Variable(tf.random_normal([n_hidden_1])), 
    'out': tf.Variable(tf.random_normal([n_classes]))} 

Construisons le graphique:

prediction = multilayer_perceptron(x, weights, biases, output) 

cost = tf.reduce_mean(tf.square(prediction-y)) #MSE 
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost) #Gives the smallest cost 

init = tf.initialize_all_variables() 

Enfin, nous allons lancer la session:

with tf.Session() as Sess: 
    Sess.run(init) 
    for epoch in range(training_epochs): 
     avg_cost = 0. 
     number_of_bathces = len(X_train)/batch_size  
     _, c = Sess.run([optimizer, cost], feed_dict = {x: X_train, y: y_train}) 
     avg_cost += c/len(X_train) 
     print(Sess.run(output_0)) 
     if epoch % display_step ==0: 
      print("Epoch:", '%02d' % (epoch+1), "cost =", "{:.9f}".format(avg_cost)) 
    print("Optimization finished") 
    correct_prediction = tf.equal(tf.arg_max(prediction,1), tf.arg_max(y,1)) 
    accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float")) 
    print("Accuracy:", accuracy.eval({x:X_test, y:y_test})) 

Et maintenant, nous obtenons la sortie:

[[ 0.77995574]] 
Epoch: 01 cost = 262.544189453 
Optimization finished 
Accuracy: 1.0 

La chose la plus déroutant est le sortie (premier nombre)! Il devrait être quelque part dans la gamme de [30; 50]! S'il vous plaît, expliquez-moi, où ai-je fait mal.

Answer 1

Votre code est notamment en désordre, donc je l'ai retiré beaucoup de pièces redondantes:

from __future__ import print_function 
import numpy as np 
import tensorflow as tf 

X_train = np.array([[ 10.], [ 10.], [ 11.], [ 6.], [ 8.], [ 9.], [ 22.], [ 14.], [ 6.], [ 8.], [ 11.], [ 9.], [ 13.], [ 7.], [ 13.], [ 7.], [ 13.], [ 11.]]) 

y_train = np.array([[ 44.5825], [ 53.99 ], [ 52.4475], [ 37.6 ], [ 38.6125], [ 39.5875], [ 43.07 ], [ 74.8575], [ 34.185 ], [ 38.61 ], [ 34.8175], [ 36.61 ], [ 34.0675], [ 37.67 ], [ 49.725 ], [ 79.4775], [ 50.41 ], [ 51.26 ]]) 

X_test = np.array([[ 6.], [ 14.], [ 14.], [ 12.], [ 13.], [ 13.]]) 

y_test = np.array([[ 55.75 ], [ 33.035 ], [ 38.3275], [ 39.2825], [ 50.7325], [ 45.2575]]) 

learning_rate = 0.05 
training_epochs = 10 

n_classes = 1 
n_hidden_1 = 5 
n_hidden_2 = 5 
n_input = 1 
x = tf.placeholder(tf.float32, [None, 1]) 
y = tf.placeholder(tf.float32, [None, 1]) 

def multilayer_perceptron(x, weights, biases): 
    layer_1 = tf.add(tf.matmul(x, weights['h1']), biases['b1']) 
    out_layer = tf.add(tf.matmul(layer_1, weights['out']), biases['out']) 
    return out_layer 

weights = { 
    'h1': tf.Variable(tf.random_normal([n_input, n_hidden_1])), 
    'out': tf.Variable(tf.random_normal([n_hidden_2, n_classes]))} 
biases = { 
    'b1': tf.Variable(tf.random_normal([n_hidden_1])), 
    'out': tf.Variable(tf.random_normal([n_classes]))} 

prediction = multilayer_perceptron(x, weights, biases) 

cost = tf.reduce_mean(tf.square(prediction - y)) #MSE 
optimizer = tf.train.AdamOptimizer(learning_rate = learning_rate).minimize(cost) #Gives the smallest cost 

init = tf.initialize_all_variables() 

with tf.Session() as sess: 
    sess.run(init) 
    for epoch in range(training_epochs): 
     _, c = sess.run([optimizer, cost], feed_dict = {x: X_train, y: y_train}) 
     print("Epoch:", '%02d' % (epoch+1), "cost =", "{:.9f}".format(c)) 
    print("Optimization finished") 

    print(sess.run(prediction, feed_dict = {x: X_test, y: y_test})) 

Il semble fonctionner maintenant. J'ai les résultats suivants:

Epoch: 01 cost = 1323.519653320 
Epoch: 02 cost = 926.386840820 
Epoch: 03 cost = 628.072326660 
Epoch: 04 cost = 431.689270020 
Epoch: 05 cost = 343.259063721 
Epoch: 06 cost = 355.978668213 
Epoch: 07 cost = 430.280548096 
Epoch: 08 cost = 501.149414062 
Epoch: 09 cost = 527.575683594 
Epoch: 10 cost = 507.708007812 
Optimization finished 
[[ 30.79703712] 
[ 69.70319366] 
[ 69.70319366] 
[ 59.97665405] 
[ 64.83992004] 
[ 64.83992004]] 

Les résultats peuvent varier en raison de l'initialisation aléatoire des poids.

Quelques conseils:

1. Utiliser plus petit taux d'apprentissage
2. train sur plusieurs époques pour voir la dynamique