Tensorflow - How to perform tf.gather with batch dimension

user_12615609 0
Viewed 336

Unfortunately I dont know how to formulate the title of this questions, maybe someone can change it?

How to replace the following for loop in an elegant way?

#tensor.shape -> (batchsize,100)
#indices.shape -> (batchsize,100) 
liste = []
for i in range(tensor.shape[0]):
        liste.append(tf.gather(tensor[i,:], indices[i,:10]))

new_tensor = tf.stack(liste)
tensorflow python
1 Answers

This should do the trick:

new_tensor = tf.gather(tensor, axis=-1, indices=indices[:, :10], batch_dims=1)

Here with a minimal reproducible example:

import tensorflow as tf

# for version 1.x
#tf.enable_eager_execution()

tensor = tf.random.normal((2, 10))
indices = tf.random.uniform(shape=[2, 10], minval=0, maxval=4, dtype=tf.int32)

liste = []
for i in range(tensor.shape[0]):
        liste.append(tf.gather(tensor[i,:], indices[i,:5]))

new_tensor = tf.stack(liste)

print('tensor: ')
print(tensor)

print('new_tensor: ')
print(new_tensor)

new_tensor_v2 = tf.gather(tensor, axis=-1, indices=indices[:, :5], batch_dims=1)
print('new_tensor_v2: ')
print(new_tensor_v2)
user_57785890
Related