numpy's hstack on tensorflow for a single matrix/tensor

Question

The numpy version of hstack for a single matrix

c=np.array([[[2,3,4],[4,5,6]],[[20,30,40],[40,50,60]]])

np.hstack(c)

output:

array([[ 2,  3,  4, 20, 30, 40],
       [ 4,  5,  6, 40, 50, 60]])

I am hoping to achieve the same behavior in TF.

c_t=tf.constant(c)
tf.stack(c_t,axis=1).eval()

I am getting the error

TypeError: Expected list for 'values' argument to 'pack' Op, not <tf.Tensor 'Const_14:0' shape=(2, 2, 3) dtype=int64>.

So I tried

tf.stack([c_t],axis=1).eval()

The output

array([[[[ 2,  3,  4],
         [ 4,  5,  6]]],


       [[[20, 30, 40],
         [40, 50, 60]]]])

I am not looking for the behaviour. tf.reshape and tf.concat are not helping me either.

`np.hstack` treats `c` as a list of 2 (2,3) arrays. In fact it starts with a list comprehension on the argument, `[atleast_1d(_m) for _m in tup]` — hpaulj, Dec 01 '18 at 18:00

score 2 · Answer 1 · answered Dec 01 '18 at 15:43

2

We can swap/permute axes and reshape -

tf.reshape(tf.transpose(c_t,(1,0,2)),(c_t.shape[1],-1))

Relevant - Intuition and idea behind reshaping 4D array to 2D array in NumPy

answered Dec 01 '18 at 15:43

Divakar

218,885
19
262
358

A programmer needs to keep track of so many axes. I will accept the previous answer for simplicity. – Proton Boss Dec 01 '18 at 15:44

score 1 · Accepted Answer · answered Dec 01 '18 at 15:37

1

One way to make it work is first unstack the tensor into a list, and then concatenate the tensors in list on first axis:

new_c = tf.concat(tf.unstack(c_t), axis=1)
sess.run(new_c)

array([[ 2,  3,  4, 20, 30, 40],
       [ 4,  5,  6, 40, 50, 60]])

answered Dec 01 '18 at 15:37

Gerges

6,269
2
22
44

Split into two tensors and concat. Nice answer! – Proton Boss Dec 01 '18 at 15:42
You should mention caveat with `tf.unstack`. It will not accept dynamic size. I can not give matrix of size (-1,2,3). -1 will throw `tf.unstack` off. – Proton Boss Dec 01 '18 at 23:23

score 1 · Answer 3 · answered Dec 02 '18 at 02:35

If you want to do it the manual way at the atomic level, then the below approach would as well work.

In [132]: c=np.array([[[2,3,4],[4,5,6]],[[20,30,40],[40,50,60]]])
In [133]: tfc = tf.convert_to_tensor(c) 

In [134]: slices = [tf.squeeze(tfc[:1, ...]), tf.squeeze(tfc[1:, ...])]  
In [135]: stacked = tf.concat(slices, axis=1) 
In [136]: stacked.eval()           
Out[136]: 
array([[ 2,  3,  4, 20, 30, 40],
       [ 4,  5,  6, 40, 50, 60]])

numpy's hstack on tensorflow for a single matrix/tensor

3 Answers3