python - TensorFlow for binary classification

Question

I am trying to adapt this MNIST example to binary classification.

But when changing my NLABELS from NLABELS=2 to NLABELS=1, the loss function always returns 0 (and accuracy 1).

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

from tensorflow.examples.tutorials.mnist import input_data
import tensorflow as tf

# Import data
mnist = input_data.read_data_sets('data', one_hot=True)
NLABELS = 2

sess = tf.InteractiveSession()

# Create the model
x = tf.placeholder(tf.float32, [None, 784], name='x-input')
W = tf.Variable(tf.zeros([784, NLABELS]), name='weights')
b = tf.Variable(tf.zeros([NLABELS], name='bias'))

y = tf.nn.softmax(tf.matmul(x, W) + b)

# Add summary ops to collect data
_ = tf.histogram_summary('weights', W)
_ = tf.histogram_summary('biases', b)
_ = tf.histogram_summary('y', y)

# Define loss and optimizer
y_ = tf.placeholder(tf.float32, [None, NLABELS], name='y-input')

# More name scopes will clean up the graph representation
with tf.name_scope('cross_entropy'):
    cross_entropy = -tf.reduce_mean(y_ * tf.log(y))
    _ = tf.scalar_summary('cross entropy', cross_entropy)
with tf.name_scope('train'):
    train_step = tf.train.GradientDescentOptimizer(10.).minimize(cross_entropy)

with tf.name_scope('test'):
    correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
    accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
    _ = tf.scalar_summary('accuracy', accuracy)

# Merge all the summaries and write them out to /tmp/mnist_logs
merged = tf.merge_all_summaries()
writer = tf.train.SummaryWriter('logs', sess.graph_def)
tf.initialize_all_variables().run()

# Train the model, and feed in test data and record summaries every 10 steps

for i in range(1000):
    if i % 10 == 0:  # Record summary data and the accuracy
        labels = mnist.test.labels[:, 0:NLABELS]
        feed = {x: mnist.test.images, y_: labels}

        result = sess.run([merged, accuracy, cross_entropy], feed_dict=feed)
        summary_str = result[0]
        acc = result[1]
        loss = result[2]
        writer.add_summary(summary_str, i)
        print('Accuracy at step %s: %s - loss: %f' % (i, acc, loss)) 
   else:
        batch_xs, batch_ys = mnist.train.next_batch(100)
        batch_ys = batch_ys[:, 0:NLABELS]
        feed = {x: batch_xs, y_: batch_ys}
    sess.run(train_step, feed_dict=feed)

I have checked the dimensions of both batch_ys (fed into y) and _y and they are both 1xN matrices when NLABELS=1 so the problem seems to be prior to that. Maybe something to do with the matrix multiplication?

I actually have got this same problem in a real project, so any help would be appreciated... Thanks!

Answer 1

The original MNIST example uses a one-hot encoding to represent the labels in the data: this means that if there are NLABELS = 10 classes (as in MNIST), the target output is [1 0 0 0 0 0 0 0 0 0] for class 0, [0 1 0 0 0 0 0 0 0 0] for class 1, etc. The tf.nn.softmax() operator converts the logits computed by tf.matmul(x, W) + b into a probability distribution across the different output classes, which is then compared to the fed-in value for y_.

If NLABELS = 1, this acts as if there were only a single class, and the tf.nn.softmax() op would compute a probability of 1.0 for that class, leading to a cross-entropy of 0.0, since tf.log(1.0) is 0.0 for all of the examples.

There are (at least) two approaches you could try for binary classification:

1. The simplest would be to set NLABELS = 2 for the two possible classes, and encode your training data as [1 0] for label 0 and [0 1] for label 1. This answer has a suggestion for how to do that.

2. You could keep the labels as integers 0 and 1 and use tf.nn.sparse_softmax_cross_entropy_with_logits(), as suggested in this answer.