Home > Software engineering >  Pad the last dimension of a tensor with different lengths to a specific length
Pad the last dimension of a tensor with different lengths to a specific length

Time:08-01

I have a similar question, this one TensorFlow - Pad unknown size tensor to a specific size?. My question is more difficult though, and I didn't find any solutions can solve my question. My question is that what if the given unknown tensor have different sizes in the last dimension and I want to pad them to the same fix length, how can I do that? For example, suppose the given tensor is

[[1],
 [1, 2],
 [1, 2, 3]]

I want to pad them such that I can get

[[1, 0, 0, 0],
 [1, 2, 0, 0],
 [1, 2, 3, 0]]

The solutions in the original post all assume the last dimension have the same length. Any ideas on how to solve this problem? I am not even sure if tf.pad() is the right function to achieve this...

CodePudding user response:

Have a look at pad_sequences

It works as follow:

sequence = [
 [1],
 [1, 2],
 [1, 2, 3]
]

tf.keras.preprocessing.sequence.pad_sequences(sequence, padding='post')

Should give you:

array([
  [1, 0, 0]
  [1, 2, 0]
  [1, 2, 3]
])

CodePudding user response:

Try combining tf.slicetf.pad and tf.map_fn.

  • For TF1
"""
[
    [1],
    [1, 2],
    [1, 2, 3]
]
"""
a = tf.sparse.SparseTensor(
    indices=[[0,0], [1,0], [1,1], [2,0], [2,1], [2,2]],
    values=[1, 1, 2, 1, 2, 3],
    dense_shape=[3, 3],
)

def cut_or_pad_1d(lst, max_len):
    origin_len = tf.shape(lst)[0]
    # cut
    lst = tf.cond(origin_len > max_len,
                  true_fn=lambda: lst[:max_len],
                  false_fn=lambda: lst)
    # pad
    lst = tf.cond(origin_len < max_len,
                  true_fn=lambda: tf.pad(lst, [[0, max_len-origin_len]]),
                  false_fn=lambda: lst)
    return lst
    
sess = tf.Session()
a_dense = tf.sparse.to_dense(a)
import functools
for MAX_LEN in (2, 5):
    a_regularized = tf.map_fn(functools.partial(cut_or_pad_1d, max_len=MAX_LEN), a_dense)
    a_regularized_val = sess.run(a_regularized)
    print(f'max_len={MAX_LEN}, a_regularized_val=')
    print(a_regularized_val)
  • For TF2
"""
[
    [1],
    [1, 2],
    [1, 2, 3]
]
"""
a = tf.sparse.SparseTensor(
    indices=[[0,0], [1,0], [1,1], [2,0], [2,1], [2,2]],
    values=[1, 1, 2, 1, 2, 3],
    dense_shape=[3, 3],
)

def cut_or_pad_1d(lst, max_len):
    origin_len = tf.shape(lst)[0]
    if origin_len > max_len:
        # cut
        lst = lst[:max_len]
    elif origin_len < max_len:
        # pad
        lst = tf.pad(lst, [[0, max_len-origin_len]])
    return lst
    
a_dense = tf.sparse.to_dense(a)
import functools
for MAX_LEN in (2, 5):
    a_regularized = tf.map_fn(functools.partial(cut_or_pad_1d, max_len=MAX_LEN), a_dense)    
    print(f'max_len={MAX_LEN}, a_regularized_val=')
    print(a_regularized.numpy())
  • Related