CODEWITHSUNDEEP
pythontensorflow
AI_ROBOT
AI_ROBOT

Reputation: 1008

Why tensorflow equal gives incorrect result

The tf.equal() function seems to be weired when we apply a CNN network. In the case below, tf.equal() returns incorrect result. 

with tf.Graph().as_default():

            images, labels = inputs("./test_data", [64, 64], 10, True)

            logits = inference(images, 2, 1.0)

            acc = accuracy(logits, labels)

            saver = tf.train.Saver()

            #predict_image(saver, logits)

            eval_once(saver, logits, acc, labels)



def eval_once(saver, logits, acc, labels):
    with tf.Session() as sess:
            ckpt = tf.train.get_checkpoint_state("./model/")
            if ckpt and ckpt.model_checkpoint_path:
                    saver.restore(sess, ckpt.model_checkpoint_path)
                    print "Model Loaded!"
            else:
                    print "Model Not Found!"
                    return

            coord = tf.train.Coordinator()

            threads = tf.train.start_queue_runners(sess, coord = coord)

            l = tf.argmax(labels,1)
            p = tf.argmax(logits,1)

            print "labels"
            print sess.run(l)

            print "preds"
            print sess.run(p)

            print sess.run(tf.equal(l, p))

            print "%.5f" % sess.run(acc)

            coord.request_stop()

            coord.join(threads, stop_grace_period_secs = 10)

In the code, inputs would read the images by file_queue and inference defines our CNN network. The output of inference is the logits of the last fully connected layer. In my case, the label is one_hot encoded and there are 2 classes, so it should be [1,0] or [0,1].

The result is below:

labels
[0 0 1 0 1 0 0 0 1 0]
preds
[0 1 0 1 0 0 1 0 1 1]
[ True  True  True  True  True  True  True  True  True  True]
0.90000

From the result we find out that labels and preds are not equal at index 1 (start from index 0). However, the tf.equal() gives us True. It is as same as index 2 and index 3.

Then I test tf.equal() on some other cases and the result is correct.

So, how could it happen?

(I used a file_queue to read the images and apply tf.train.batch or tf.train.shuffle to create the batch for training and testing. Just like the example of cifar 10 in tensorflow examples)

Upvotes: 0

Views: 381

Answers (1)

AI_ROBOT
AI_ROBOT

Reputation: 1008

After the thinking, there is one possibility that each time I call sess.run() the file_queue would read shuffled images and labels so that the result is weired.

Upvotes: 1

Related Questions