tensorflow一个奇怪的问题

Question

这是《Machine Learning with Tensorflow》上的一个例程，它使用了batch如下

    for step in range(training_epochs * train_size // batch_size):
        offset = (step * batch_size) % train_size
        batch_xs = xs[offset:(offset + batch_size), :]
        batch_labels = labels[offset:(offset + batch_size)]
        err, _ = sess.run([cost, train_op], feed_dict={X: batch_xs, Y: batch_labels})
        if step % 100 == 0:
                    print (step, err)

此时程序正常运行，但如果把这一段改成:

    for epoch in range(training_epochs):
        err, _ = sess.run([cost, train_op], feed_dict={X:xs,Y:labels})
        if(epoch%100==0):
            print(epoch,err)

逻辑上没问题，但运行结果都是nan。这是怎么回事？

下面是完整程序

import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt

learning_rate = 0.01
training_epochs = 1000
num_labels = 3
batch_size = 100

x1_label0 = np.random.normal(1, 1, (100, 1))
x2_label0 = np.random.normal(1, 1, (100, 1))
x1_label1 = np.random.normal(5, 1, (100, 1))
x2_label1 = np.random.normal(4, 1, (100, 1))
x1_label2 = np.random.normal(8, 1, (100, 1))
x2_label2 = np.random.normal(0, 1, (100, 1))

xs_label0 = np.hstack((x1_label0, x2_label0))
xs_label1 = np.hstack((x1_label1, x2_label1))
xs_label2 = np.hstack((x1_label2, x2_label2))

xs = np.vstack((xs_label0, xs_label1, xs_label2))
labels = np.matrix([[1., 0., 0.]] * len(x1_label0) + [[0., 1., 0.]] * len(x1_label1) + [[0., 0., 1.]] * len(x1_label2))

arr = np.arange(xs.shape[0])
np.random.shuffle(arr)
xs = xs[arr, :]
labels = labels[arr, :]


train_size, num_features = xs.shape

X = tf.placeholder("float", shape=[None, num_features])
Y = tf.placeholder("float", shape=[None, num_labels])

W = tf.Variable(tf.zeros([num_features, num_labels]))
b = tf.Variable(tf.zeros([num_labels]))
y_model = tf.nn.softmax(tf.matmul(X, W) + b)

cost = -tf.reduce_sum(Y * tf.log(y_model))
train_op = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)

correct_prediction = tf.equal(tf.argmax(y_model, 1), tf.argmax(Y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))

with tf.Session() as sess:
    tf.initialize_all_variables().run()

#    for step in range(training_epochs * train_size // batch_size):
#        offset = (step * batch_size) % train_size
#        batch_xs = xs[offset:(offset + batch_size), :]
#        batch_labels = labels[offset:(offset + batch_size)]
#        err, _ = sess.run([cost, train_op], feed_dict={X: batch_xs, Y: batch_labels})
#        if step % 100 == 0:
#            print (step, err)
    for epoch in range(training_epochs):
        err, _ = sess.run([cost, train_op], feed_dict={X:xs,Y:labels})
        if(epoch%100==0):
            print(epoch,err)

    W_val = sess.run(W)
    print('w', W_val)
    b_val = sess.run(b)
    print('b', b_val)

Answer 1

不知道你这个问题是否已经解决, 如果还没有解决的话:

• 你可以参考下这篇文章：Tensorflow计算一个模型的浮点运算数

如果你已经解决了该问题, 非常希望你能够分享一下解决方案, 以帮助更多的人 ^-^