CODEWITHSUNDEEP
pythontensorflow
User
User

Reputation: 826

Add a tensor only to a part of another tensor

I have to add two tensors, one with a shape multiple of the other in the depth direction. Here an example

t1 = tf.constant(3, shape=[2, 2, 2], dtype=tf.float32)
t2 = tf.constant(1, shape=[2, 2, 1], dtype=tf.float32)

I want to use something like tf.add to add the second tensor to the first but only in the first layer of the third component of the shape. With numbers

t1 = [[[3, 3], [3, 3]],
      [[3, 3], [3, 3]]]
t2 = [[[1, 1], [1, 1]]]

output = [[[4, 4], [4, 4]],
          [[3, 3], [3, 3]]]

Is there a built-in function to do that?

Upvotes: 1

Views: 295

Answers (1)

akuiper
akuiper

Reputation: 214927

Add the first 'column' of t1 with t2 and then concat it with the rest columns of t1:

t1 = tf.constant(3, shape=[2, 2, 2], dtype=tf.float32)
t2 = tf.constant(1, shape=[2, 2, 1], dtype=tf.float32)
tf.InteractiveSession()

tf.concat((t1[...,0:1] + t2, t1[...,1:]), axis=2).eval()

#array([[[4., 3.],
#        [4., 3.]],

#       [[4., 3.],
#        [4., 3.]]], dtype=float32)

Notice your second example t2 has a different shape, i.e. (1,2,2) instead of (2,2,1), in which case, slice and concat by the first axis:

tf.concat((t1[0:1] + t2, t1[1:]), axis=0).eval()

#array([[[4., 4.],
#        [4., 4.]],

#       [[3., 3.],
#        [3., 3.]]], dtype=float32)

Upvotes: 2

Related Questions