[tensorflow] which dimension to reduce in tf.reduce_sum()

I began to learn tensorflow.

import tensorflow as tf

, firstly of course. When encountered with

tf.reduce_sum(array)

or

tf.reduce_mean(array)

, I felt a little doubt about which is the exact dimension to be reduced. Directly digging the answer in the unfamiliar tensorflow world(for me), somehow, made me confused.

the answer

Think about this. When create a array applying numpy, we use

a = numpy.random.rand(3,4,5)

. That means the length of the first dimension is 3, the length of the second dimension is 4 and 5 for the third dimension.

if applying

sess.run(tf.reduce_sum(a,0))

, the product’s shape will be

(4,5)

, the dimension length of which is 3 is reduced.

appendix

import tensorflow as tf
import numpy as np
a = np.random.rand(3,4,5)
a
Out[138]:
array([[[ 0.67386161,  0.25891236,  0.15750355,  0.99356577,  0.62401749],
[ 0.44104193,  0.14171052,  0.92457023,  0.3475105 ,  0.46261946],
[ 0.74483249,  0.22724867,  0.61261517,  0.20140201,  0.11718528],
[ 0.53023319,  0.65372313,  0.34679634,  0.7626164 ,  0.47658279]],
[[ 0.22209199,  0.57104615,  0.94053357,  0.10663142,  0.96630193],
[ 0.87147539,  0.29464845,  0.41552753,  0.05044025,  0.92632825],
[ 0.78404338,  0.42560083,  0.91265402,  0.37281405,  0.25450812],
[ 0.00306304,  0.74638202,  0.19689413,  0.65906257,  0.46627029]],
[[ 0.46042323,  0.48506186,  0.73388123,  0.50179246,  0.3163692 ],
[ 0.33435115,  0.01610695,  0.98188888,  0.77100164,  0.7795511 ],
[ 0.24383665,  0.28206927,  0.09408851,  0.90500411,  0.69718288],
[ 0.40164087,  0.66995977,  0.61219998,  0.91530942,  0.00388272]]])
sess = tf.Session()
sess.run(tf.reduce_mean(a,0))
Out[140]:
array([[ 0.45212561,  0.43834012,  0.61063945,  0.53399655,  0.63556287],
[ 0.54895615,  0.15082197,  0.77399555,  0.3896508 ,  0.72283294],
[ 0.59090417,  0.31163959,  0.5397859 ,  0.49307339,  0.35629209],
[ 0.3116457 ,  0.69002164,  0.38529681,  0.77899613,  0.3155786 ]])
sess.run(tf.reduce_mean(a,0)).shape
Out[141]: (4, 5)
sess.run(tf.reduce_mean(a,1)).shape
Out[142]: (3, 5)
In[...]: import tensorflow as tf
In[...]: import numpy as np
In[...]: a = np.random.rand(3,4,5)
In[...]: a
Out[138]:
array([[[ 0.67386161,  0.25891236,  0.15750355,  0.99356577,  0.62401749],
[ 0.44104193,  0.14171052,  0.92457023,  0.3475105 ,  0.46261946],
[ 0.74483249,  0.22724867,  0.61261517,  0.20140201,  0.11718528],
[ 0.53023319,  0.65372313,  0.34679634,  0.7626164 ,  0.47658279]],
[[ 0.22209199,  0.57104615,  0.94053357,  0.10663142,  0.96630193],
[ 0.87147539,  0.29464845,  0.41552753,  0.05044025,  0.92632825],
[ 0.78404338,  0.42560083,  0.91265402,  0.37281405,  0.25450812],
[ 0.00306304,  0.74638202,  0.19689413,  0.65906257,  0.46627029]],
[[ 0.46042323,  0.48506186,  0.7
9d49
3388123,  0.50179246,  0.3163692 ],
[ 0.33435115,  0.01610695,  0.98188888,  0.77100164,  0.7795511 ],
[ 0.24383665,  0.28206927,  0.09408851,  0.90500411,  0.69718288],
[ 0.40164087,  0.66995977,  0.61219998,  0.91530942,  0.00388272]]])
In[...]: sess = tf.Session()
In[...]: sess.run(tf.reduce_mean(a,0))
Out[140]:
array([[ 0.45212561,  0.43834012,  0.61063945,  0.53399655,  0.63556287],
[ 0.54895615,  0.15082197,  0.77399555,  0.3896508 ,  0.72283294],
[ 0.59090417,  0.31163959,  0.5397859 ,  0.49307339,  0.35629209],
[ 0.3116457 ,  0.69002164,  0.38529681,  0.77899613,  0.3155786 ]])
In[...]: sess.run(tf.reduce_mean(a,0)).shape
Out[141]: (4, 5)
In[...]: sess.run(tf.reduce_mean(a,1)).shape
Out[142]: (3, 5)