TensorFlow 中是否有与 python reduce 等效的函数？

Question

I want a function in TensorFlow which has the same effect as Pythion reduce()

For example, if I have a tensor a with value [a1, a2, a3] and a function func() , I want [func(a1, a2), func(func(a1, a2), a3)] . If a is a Python list I can simply do

from functools import reduce
# a = [a1, a2, a3]
A = reduce(func, a)

What if a is a Tensor (not a python list!) of TF object and I want the same A? If no alternative TF function, how can I do it effectively?

Answer 1

If you actually mean reduce, so the tensorflow equivalent are tf.foldl and tf.foldr . Example:

elems = tf.constant(["hello", "my", "name", "is", "inigo", "montoya"])
tf.foldl(lambda a, x: a +" " + x, elems)
#<tf.Tensor: shape=(), dtype=string, numpy=b'hello my name is inigo montoya'>
tf.foldr(lambda a, x: a +" " + x, elems)
#<tf.Tensor: shape=(), dtype=string, numpy=b'montoya inigo is name my hello'>

However, according to your example you probably want tf.scan :

tf.scan(lambda a, x: a +" " + x, elems)
<tf.Tensor: shape=(6,), dtype=string, numpy=
    ([b'hello', b'hello my', b'hello my name', b'hello my name is',
    b'hello my name is inigo', b'hello my name is inigo montoya'], dtype=object)>

Answer 2

You have not mentioned which type of func you would like to achieve, ie whether you just wanted to simply add all tensors, or take mean etc. For those type of operations there is tf.reduce_sum , and set of other reduce operations. But if you want to have your own custom operation to be implemented by means of reduce, you can achieve it with tf.py_function . Check the code below:

import tensorflow as tf
from functools import reduce

tf.enable_eager_execution()

x = tf.convert_to_tensor(tf.random.normal([1, 4]))
y = tf.convert_to_tensor(tf.random.normal([1, 4]))
z = tf.convert_to_tensor(tf.random.normal([1, 4]))

print('X = ', x.numpy())
print('Y = ', y.numpy())
print('Z = ', z.numpy())

def custom_function(*elements):
    data = reduce(lambda x, y: x + y, elements)
    return data


result = tf.py_function(custom_function, [x, y, z], Tout=tf.float32)
print('Result = ', result.numpy())

Output:

X =  [[ 0.9403639  -0.41230875  0.08281742  0.24365784]]
Y =  [[-1.2166423  -0.18330204 -1.2322044  -1.1254508 ]]
Z =  [[-0.42474216  0.997245   -2.319666    1.4623599 ]]
Result =  [[-0.70102054  0.40163422 -3.4690528   0.5805669 ]]

EDIT:

The output of any reduce function in Python has to be a single value. In your case, you wish to accumulate the values of subsequent elements in a tensor and in the end produce another 1 dimensional Tensor.

I can't think of any built-in TensorFlow equivalent operation which can be used in this way. But, we can do the processing by indexing of tensor elements. However, I am not sure, how much this solution can integrate with your overall model architecture. Check the code below:

import tensorflow as tf
tf.enable_eager_execution()

x = tf.convert_to_tensor(tf.random.normal([5]))
print('X = ', x.numpy())

y = []
sum = x[0]
for index in range(1, len(x)):
    sum += x[index]
    y.append(sum)
y = tf.convert_to_tensor(y)
print('Sum = ', y.numpy())

Output:

X =  [ 1.0281714  -1.5870513  -1.7896062   0.11207424  0.50049555]
Sum =  [-0.55887985 -2.348486   -2.2364118  -1.7359163 ]