Tensorflow error: Minimum tensor rank: 1 but got: 1

Question

I'm getting the following error trying to eval my model.

tensorflow.python.framework.errors.InvalidArgumentError: Minimum tensor rank: 1 but got: 1 [[Node: ArgMax_1 = ArgMax[T=DT_INT64, _device="/job:localhost/replica:0/task:0/cpu:0"](_recv_Placeholder_1_0, ArgMax_1/dimension/_40)]]

Here is the relevant code

        # Predictions for the current training minibatch.
        train_prediction = tf.nn.softmax(logits)
        correct_prediction = tf.equal(tf.argmax(train_prediction, 1), tf.argmax(train_labels, 1))
        accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

        sess.run(tf.initialize_all_variables())
        for i in range(1000000):
            start_time = time()

            images, labels = get_batch(fifo_queue, FLAGS.batch_size)

            feed_dict = {
                train_images: images,
                train_labels: labels
            }

            _, loss_value, learn_rate, predictions = sess.run(
                [train_step, cross_entropy, learning_rate, train_prediction],
                feed_dict=feed_dict)

            duration = time() - start_time
              if i % 1 == 0:
                # Print status to stdout.
                 print('Step %d: loss = %.3f (%.3f sec)' % (i, loss_value, duration))

                 train_accuracy = accuracy.eval(feed_dict={
                     train_images: images, train_labels: labels, keep_prob: 1.0})
                 print("step %d, training accuracy %g"%(i, train_accuracy))
                 train_step.run(feed_dict={train_images: images[0], train_labels: labels[1], keep_prob: 0.5})

`

I haven't been able to try much yet because I'm just getting my first model eval-ing and this error (indicating expecting 1 and got 1) is not overly helpful.

From what I see it seems fine, provided that the function get_batch() returns tensors with the right shapes. Try to take a look at this example, and compare the rest of the code (variales and placeholders definition, data tensor shapes, etc). — Diogo Pinto
This code runs except for the eval part. Argmax is supposed to take the predictions and the actual labels or the label placeholder, it seems. Wonder why this isn't working... — JohnAllen
Just a thing, shouldn't the argument to argmax train_labels_node be the same that you use in your feed_dict (where you have simply train_labels? — Diogo Pinto
I'll try to clarify argmax, it may help: as you are using, what arrives at each invocation should be a 2D matrix, each row being a sample, and each column being the probability that a sample is of the class ith-column (train labels should have just one column set to 1 for each sample, and train_prediction a distribution of probabilities among the classes). So, equal is comparing, for each sample, the index of the column with the highest value, returning a boolean column. — Diogo Pinto
Yes, for the setup you have you should use one-hot encoding. Argmax does nothing in there. Otherwise you would have to change your accuracy function (to use a threshold on the predictions, from that create a vector of zeros and ones, and now you could apply directly the tf.equals. — Diogo Pinto

mrry mrry · Accepted Answer · 2016-06-03T21:51:00

The error message isn't great, but looking at the code might explain what's going on.

The issue arises because train_labels is (presumably) a one-dimensional vector. Dimensions are numbered from 0, so a vector only has a 0th dimension, but your invocation of tf.argmax(train_labels, 1) attempts to take the argmax in the 1st dimension, which doesn't exist.

In fact, there's no need to take the argmax of the labels at all. Instead, you can simply write:

correct_prediction = tf.equal(tf.argmax(train_prediction, 1), train_labels)

Tensorflow error: Minimum tensor rank: 1 but got: 1

1 Answers