tf.concat Function

Just how to use tf.concat Function, that function concatenates tensor along one dimension you specified like this.

tf.concat(list, dimension you want to concatenate). i.e.

If tensors, values[i].shape is [D0, D1, …, Daxis(i), …, Dn], the resulting concatenation’s shape is

[D0, D1, …, Raxis(concatenated), …, Dn]

In the case of the function,

Raxis = sum(Daxis(i))

So That is the data from input tensors is joined along the axis dimension.

Be careful of this following:

The number of dimensions of the input tensors must match and all dimension except axis must be equal

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

"""Example code for tf.concat(https://www.tensorflow.org/versions/r1.8/api_docs/python/tf/concat)

tf.concat(
    values,
    axis,
    name='concat'
)
"""

import sys
import tensorflow as tf

print("=== Version checking ===")
print("The version of sys: \n{}".format(sys.version))
print("Tensorflow version: {}".format(tf.__version__))
print("========================")

=== Version checking ===
The version of sys: 
3.5.2 (default, Nov 23 2017, 16:37:01) 
[GCC 5.4.0 20160609]
Tensorflow version: 1.8.0
========================
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18

test1 = tf.constant([[1, 2, 3], [4, 5, 6]])
test2 = tf.constant([[7, 8, 9], [10, 11, 12]])

concat1 = tf.concat([test1, test2], 0)
concat2 = tf.concat([test1, test2], 1)

# Also, you can use negative index, the order is opposite to the correct order. 
# I mean, concat1 is the same from concat1_1, and concat2 is the same from concat2_2
concat2_1 = tf.concat([test1, test2], -2)
concat2_2 = tf.concat([test1, test2], -1)

print("===== Tensor Shape ======")
print("test1: {}".format(test1))
print("test2: {}".format(test2))
print("concat1 with D-0: {}".format(concat1))
print("concat2 with D-1: {}".format(concat2))
print("concat2_1 with D-0(-2): {}".format(concat2_1))
print("concat2_2 with D-1(-1): {}".format(concat2_2))

===== Tensor Shape ======
test1: Tensor("Const:0", shape=(2, 3), dtype=int32)
test2: Tensor("Const_1:0", shape=(2, 3), dtype=int32)
concat1 with D-0: Tensor("concat:0", shape=(4, 3), dtype=int32)
concat2 with D-1: Tensor("concat_1:0", shape=(2, 6), dtype=int32)
concat2_1 with D-0(-2): Tensor("concat_2:0", shape=(4, 3), dtype=int32)
concat2_2 with D-1(-1): Tensor("concat_3:0", shape=(2, 6), dtype=int32)
1
2
3
4
5
6
7

with tf.Session() as sess:
    print("test1: \n{}, The shape: {}".format(sess.run(test1), test1.shape))
    print("test2: \n{}, The shape: {}".format(sess.run(test2), test2.shape))
    print("concat1 with D-0: \n{}, The shape: {}".format(sess.run(concat1), concat1.shape))
    print("concat2 with D-1: \n{}, The shape: {}".format(sess.run(concat2), concat2.shape))
    print("concat2_1 with D-0(-2): \n{}, The shape: {}".format(sess.run(concat2_1), concat2_1.shape))
    print("concat2_2 with D-1(-1): \n{}, The shape: {}".format(sess.run(concat2_2), concat2_2.shape))

with tf.Session() as sess:

    print("test1: \n{}, The shape: {}".format(sess.run(test1), test1.shape))

    print("test2: \n{}, The shape: {}".format(sess.run(test2), test2.shape))

    print("concat1 with D-0: \n{}, The shape: {}".format(sess.run(concat1), concat1.shape))

    print("concat2 with D-1: \n{}, The shape: {}".format(sess.run(concat2), concat2.shape))

    print("concat2_1 with D-0(-2): \n{}, The shape: {}".format(sess.run(concat2_1), concat2_1.shape))

    print("concat2_2 with D-1(-1): \n{}, The shape: {}".format(sess.run(concat2_2), concat2_2.shape))

test1: 
[[1 2 3]
 [4 5 6]], The shape: (2, 3)
test2: 
[[ 7  8  9]
 [10 11 12]], The shape: (2, 3)
concat1 with D-0: 
[[ 1  2  3]
 [ 4  5  6]
 [ 7  8  9]
 [10 11 12]], The shape: (4, 3)
concat2 with D-1: 
[[ 1  2  3  7  8  9]
 [ 4  5  6 10 11 12]], The shape: (2, 6)
concat2_1 with D-0(-2): 
[[ 1  2  3]
 [ 4  5  6]
 [ 7  8  9]
 [10 11 12]], The shape: (4, 3)
concat2_2 with D-1(-1): 
[[ 1  2  3  7  8  9]
 [ 4  5  6 10 11 12]], The shape: (2, 6)

For more example, with normal list of python.

1
2
3
4
5
6
7
8
9
10
11
12

te1 = [[[1, 2], [2, 3]], [[4, 4], [5, 3]]]
te2 = [[[7, 4], [8, 4]], [[2, 10], [15, 11]]]

conc1 = tf.concat([te1, te2], -1)


print("===== Python list ======")
print("te1: {}".format(te1))
print("te2: {}".format(te2))

print("\n===== Tensor Shape ======")
print("conc1 with D-2: {}".format(conc1))

===== Python list ======
te1: [[[1, 2], [2, 3]], [[4, 4], [5, 3]]]
te2: [[[7, 4], [8, 4]], [[2, 10], [15, 11]]]

===== Tensor Shape ======
conc1 with D-2: Tensor("concat_4:0", shape=(2, 2, 4), dtype=int32)
1
2

with tf.Session() as sess:
    print("conc1: \n{}, The shape: {}".format(sess.run(conc1), conc1.shape))

conc1: 
[[[ 1  2  7  4]
  [ 2  3  8  4]]

 [[ 4  4  2 10]
  [ 5  3 15 11]]], The shape: (2, 2, 4)

Reference