3

Is it possible to make a binary classification? Especially for pedestrian detection, whether it is or not a pedestrian. I couldn't find anything in the API or any good tutorials for this. I tried to adapt the code from the deep MNIST tutorial, that was used for multi-class classification; I made the images with pedestrians in them labeled with 1, and the negatives with 0, and used 3 channels(for colours, shouldn't be a problem right?), but the accuracy just jumps all over the place.

The code

    import dataset as input_data
    import tensorflow as tf


    def weight_variable(shape):
        initial = tf.truncated_normal(shape, stddev=0.1)
        return tf.Variable(initial)


    def bias_variable(shape):
        initial = tf.constant(0.1, shape=shape)
        return tf.Variable(initial)


    def conv2d(x, W):
        return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')


    def max_pool_2x2(x):
        return tf.nn.max_pool(x, ksize=[1, 3, 3, 1],
                              strides=[1, 2, 2, 1], padding='SAME')


    data = input_data.read_data_sets()

    sess = tf.InteractiveSession()

    x = tf.placeholder("float", shape=[None, input_data.HEIGHT * input_data.WIDTH * 3])
    y_ = tf.placeholder("float", shape=[None, 2])

    W_conv1 = weight_variable([5, 5, 3, 64])
    b_conv1 = bias_variable([64])

    x_image = tf.reshape(x, [-1, input_data.WIDTH, input_data.HEIGHT, 3])

    h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
    h_pool1 = max_pool_2x2(h_conv1)
    h_norm1 = tf.nn.lrn(h_pool1, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75)

    W_conv2 = weight_variable([5, 5, 64, 64])
    b_conv2 = bias_variable([64])

    h_conv2 = tf.nn.relu(conv2d(h_norm1, W_conv2) + b_conv2)
    h_norm2 = tf.nn.lrn(h_conv2, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75)
    h_pool2 = max_pool_2x2(h_norm2)

    W_fc1 = weight_variable([input_data.HEIGHT / 4 * input_data.WIDTH / 4 * 64, 1024])
    b_fc1 = bias_variable([1024])

    h_pool2_flat = tf.reshape(h_pool2, [-1, input_data.HEIGHT / 4 * input_data.WIDTH / 4 * 64])
    h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)

    keep_prob = tf.placeholder("float")
    h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

    W_fc2 = weight_variable([1024, 2])
    b_fc2 = bias_variable([2])

    y_conv = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)

    cross_entropy = -tf.reduce_sum(y_ * tf.log(y_conv))
    train_step = tf.train.AdamOptimizer(1e-6).minimize(cross_entropy)
    correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))
    accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
    sess.run(tf.initialize_all_variables())
    for i in range(20000):
        batch = data.train.next_batch(50)
        if i % 100 == 0:
            train_accuracy = accuracy.eval(feed_dict={
                x: batch[0], y_: batch[1], keep_prob: 1.0})
            print "step %d, training accuracy %g" % (i, train_accuracy)
        train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})

    print "test accuracy %g" % accuracy.eval(feed_dict={
        x: data.test.images, y_: data.test.labels, keep_prob: 1.0})

The output

    step 0, training accuracy 0.14
    step 100, training accuracy 0.54
    step 200, training accuracy 0.28
    step 300, training accuracy 0.46
    step 400, training accuracy 0.32
    step 500, training accuracy 0.52
    step 600, training accuracy 0.56
    step 700, training accuracy 0.76
    step 800, training accuracy 0.66

Help would be apriciated, thanks.

Improve this question
1
  • just looking at your code, your cross_entropy = -tf.reduce_sum(y_ * tf.log(y_conv)) is not correct if you encoding 0-1 Commented Mar 16, 2017 at 18:37

1 Answer 1

Reset to default
1

You should definitely use tensorboard to visualize the cross entropy , biases and weights summaries. I think it will give you a much better view of what is going on.

Try with this code and then run tensorboard:

import dataset as input_data
import tensorflow as tf


def weight_variable(shape):
    initial = tf.truncated_normal(shape, stddev=0.1)
    return tf.Variable(initial)


def bias_variable(shape):
    initial = tf.constant(0.1, shape=shape)
    return tf.Variable(initial)


def conv2d(x, W):
    return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')


def max_pool_2x2(x):
    return tf.nn.max_pool(x, ksize=[1, 3, 3, 1],
                          strides=[1, 2, 2, 1], padding='SAME')


data = input_data.read_data_sets()

sess = tf.InteractiveSession()

x = tf.placeholder("float", shape=[None, input_data.HEIGHT * input_data.WIDTH * 3])
y_ = tf.placeholder("float", shape=[None, 2])

W_conv1 = weight_variable([5, 5, 3, 64])
b_conv1 = bias_variable([64])

x_image = tf.reshape(x, [-1, input_data.WIDTH, input_data.HEIGHT, 3])

h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
h_pool1 = max_pool_2x2(h_conv1)
h_norm1 = tf.nn.lrn(h_pool1, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75)

W_conv2 = weight_variable([5, 5, 64, 64])
b_conv2 = bias_variable([64])

h_conv2 = tf.nn.relu(conv2d(h_norm1, W_conv2) + b_conv2)
h_norm2 = tf.nn.lrn(h_conv2, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75)
h_pool2 = max_pool_2x2(h_norm2)

W_fc1 = weight_variable([input_data.HEIGHT / 4 * input_data.WIDTH / 4 * 64, 1024])
b_fc1 = bias_variable([1024])

h_pool2_flat = tf.reshape(h_pool2, [-1, input_data.HEIGHT / 4 * input_data.WIDTH / 4 * 64])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)

keep_prob = tf.placeholder("float")
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

W_fc2 = weight_variable([1024, 2])
b_fc2 = bias_variable([2])

y_conv = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)

# Add summary ops to collect data
w_fc2_hist = tf.histogram_summary("weights_fc2", W_fc2)
b_fc2_hist = tf.histogram_summary("bias_fc2", b_fc2)
w_in_hist = tf.histogram_summary("weights_in", W)
b_in_hist = tf.histogram_summary("bias_in", b)
y_hist = tf.histogram_summary("y_conv", y_conv)

cross_entropy = -tf.reduce_sum(y_ * tf.log(y_conv))
ce_summ = tf.scalar_summary("cross entropy", cross_entropy)
train_step = tf.train.AdamOptimizer(1e-6).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
accuracy_summary = tf.scalar_summary("train accuracy", accuracy)
# Merge all the summaries and write them out to /tmp/tf
merged = tf.merge_all_summaries()
writer = tf.train.SummaryWriter("/tmp/tf", sess.graph_def)
sess.run(tf.initialize_all_variables())
for i in range(20000):
    batch = data.train.next_batch(50)
    feed={x: batch[0], y_: batch[1], keep_prob: 0.5}
    result = sess.run([merged, accuracy, train_step], feed_dict=feed)        
    if i % 100 == 0:  # Record summary data, and the accuracy
            summary_str = result[0]
            acc = result[1]
            writer.add_summary(summary_str, i)
            print("Accuracy at step {0}/{1}: {2}%".format(i, 20000, int(acc*100)))

print "test accuracy %g" % accuracy.eval(feed_dict={
    x: data.test.images, y_: data.test.labels, keep_prob: 1.0})
Improve this answer
Sign up to request clarification or add additional context in comments.

5 Comments

Which weights and biases are W, b you used to add the histogram summary?
Those from the second fully connected layer, hence the "fc2" in their names. You can easily change that, or add new weights and biases and visualize all at the same time. Would you know how to do that by inspecting the code?
Well I tried the code, it converges at 100% which seems wrong. I added the summaries after I modified the code a little, I edited the post to put some screenshots with the graphs.
Graph Screenshots Cross Entropy and Mean Loss Train Accuracy(Seems to be the opposite of the loss) Bias and Weights Logits
How many images do you have for training? Your model just seems to be overfitting. You should also have a different test set of images to have a better sense of overfitting. In short, I would say there is nothing inherently wrong about your results

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.