Ytukyg

I want to calculate the Euclidean distance of the vectors in the features, which is a tf.Tensor got from the network.

I tried it in the following way, but failed with error:

'Tensor' object is not iterable

So I want to calculate the distance between the rows in one matrix just through matrix，without iteration of every rows.

features, _ = mnist_net(images)

feature_matrix = np.zeros(shape=(FLAGS.batch_size,FLAGS.batch_size))

for i in range (FLAGS.batch_size):

   for j in range (FLAGS.batch_size):

      aa = tf.slice(features,[i,0],[1,50])

      bb = tf.slice(features,[j,0],[1,50])

      feature_matrix[i,j] = tf.sqrt(sum((aa-bb)**2))  

You can achieve that simply with tf.norm/tf.linalg.norm:

feature_matrix = tf.linalg.norm(features[:, tf.newaxis] - features, axis=-1)

For example:

import tensorflow as tf



with tf.Session() as sess:

    features = tf.placeholder(tf.float32, [None, None])

    feature_matrix = tf.linalg.norm(features[:, tf.newaxis] - features, axis=-1)

    print(sess.run(feature_matrix, feed_dict={features: [[ 1,  2,  3],

                                                         [ 4,  5,  6],

                                                         [ 7,  8,  9],

                                                         [10, 11, 12]]}))

Output:

[[ 0.        5.196152 10.392304 15.588457]

 [ 5.196152  0.        5.196152 10.392304]

 [10.392304  5.196152  0.        5.196152]

 [15.588457 10.392304  5.196152  0.      ]]

EDIT:

If you cannot use tf.norm, the following is an equivalent implementation:

sqdiff = tf.squared_difference(features[:, tf.newaxis], features)

feature_matrix = tf.sqrt(tf.reduce_sum(sqdiff, axis=-1))