Files
2020-10-14 08:55:07 +08:00

46 lines
1.5 KiB
Python

import tensorflow as tf
from onnx_tf.handlers.backend_handler import BackendHandler
from onnx_tf.handlers.handler import onnx_op
@onnx_op("SequenceErase")
class SequenceErase(BackendHandler):
@classmethod
def chk_pos_in_bounds(cls, input_seq, pos):
"""
Check the position is in-bounds with respect to the sequence.
Accepted range for 'position' is in [-n, n - 1], where n is the
number of tensors in 'input_sequence'.
:param input_seq: input sequence
:param pos: position of the output tensor
:return: True if position is in-bounds
"""
seq_length = tf.shape(input_seq.to_sparse(), out_type=pos.dtype)[0]
cond1 = tf.greater_equal(pos, tf.negative(seq_length))
cond2 = tf.less_equal(pos, seq_length - 1)
# pos >= -n and pos < n
return tf.reduce_all(tf.logical_and(cond1, cond2))
@classmethod
def version_11(cls, node, **kwargs):
tensor_dict = kwargs["tensor_dict"]
input_sequence = tensor_dict[node.inputs[0]]
seq_length = tf.shape(input_sequence.to_sparse())[0]
position = tensor_dict[node.inputs[1]] if len(
node.inputs) == 2 else seq_length - 1
# check whether position is in-bounds and assert if not
result = cls.chk_pos_in_bounds(input_sequence, position)
assert_pos = tf.Assert(tf.equal(result, True), [result])
with tf.control_dependencies([assert_pos]):
s1 = input_sequence[:position]
s2 = input_sequence[position + 1:]
return [tf.concat([s1, s2], axis=0)]