import numpy as np
import tensorflow as tf
[docs]def sequence_1d(
seq, maxlen=None, padding: str = 'post', pad_int=0, return_len=False
):
"""
padding sequence of 1d to become 2d array.
Parameters
----------
seq: List[List[int]]
maxlen: int, optional (default=None)
If None, will calculate max length in the function.
padding: str, optional (default='post')
If `pre`, will add 0 on the starting side, else add 0 on the end side.
pad_int, int, optional (default=0)
padding value.
Returns
--------
result: np.array
"""
if padding not in ['post', 'pre']:
raise ValueError('padding only supported [`post`, `pre`]')
if not maxlen:
maxlen = max([len(s) for s in seq])
padded_seqs, length = [], []
for s in seq:
if isinstance(s, np.ndarray):
s = s.tolist()
if padding == 'post':
padded_seqs.append(s + [pad_int] * (maxlen - len(s)))
if padding == 'pre':
padded_seqs.append([pad_int] * (maxlen - len(s)) + s)
length.append(len(s))
if return_len:
return np.array(padded_seqs), length
return np.array(padded_seqs)
[docs]def sequence_nd(
seq,
maxlen=None,
padding: str = 'post',
pad_val=0.0,
dim: int = 1,
return_len=False,
):
"""
padding sequence of nd to become (n+1)d array.
Parameters
----------
seq: list of nd array
maxlen: int, optional (default=None)
If None, will calculate max length in the function.
padding: str, optional (default='post')
If `pre`, will add 0 on the starting side, else add 0 on the end side.
pad_val, float, optional (default=0.0)
padding value.
dim: int, optional (default=1)
Returns
--------
result: np.array
"""
if padding not in ['post', 'pre']:
raise ValueError('padding only supported [`post`, `pre`]')
if not maxlen:
maxlen = max([np.shape(s)[dim] for s in seq])
padded_seqs, length = [], []
for s in seq:
npad = [[0, 0] for _ in range(len(s.shape))]
if padding == 'pre':
padding = 0
if padding == 'post':
padding = 1
npad[dim][padding] = maxlen - s.shape[dim]
padded_seqs.append(
np.pad(
s,
pad_width=npad,
mode='constant',
constant_values=pad_val,
)
)
length.append(s.shape[dim])
if return_len:
return np.array(padded_seqs), length
return np.array(padded_seqs)
[docs]def tf_sequence_nd(
seq,
maxlen=None,
padding: str = 'post',
pad_val=0.0,
dim: int = 1,
return_len=False,
):
"""
padding sequence of nd to become (n+1)d array.
Parameters
----------
seq: list of nd array
maxlen: int, optional (default=None)
If None, will calculate max length in the function.
padding: str, optional (default='post')
If `pre`, will add 0 on the starting side, else add 0 on the end side.
pad_val, float, optional (default=0.0)
padding value.
dim: int, optional (default=1)
Returns
--------
result: np.array
"""
if not maxlen:
maxlen = tf.reduce_max([tf.shape(seq[i])[0] for i in range(len(seq))])
padded_seqs, length = [], []
for i in range(len(seq)):
s = seq[i]
npad = [[0, 0] for _ in range(len(s.shape))]
if padding == 'pre':
padding = 0
if padding == 'post':
padding = 1
npad[dim][padding] = maxlen - tf.shape(s)[dim]
padded_seqs.append(
tf.pad(
s, paddings=npad, mode='CONSTANT', constant_values=pad_val
)
)
length.append(tf.shape(s)[dim])
if return_len:
return tf.stack(padded_seqs), tf.stack(length)
return tf.stack(padded_seqs)
