ops.py File Reference

some operations More...

Functions
def	nabu.neuralnetworks.components.ops.pyramid_stack (inputs, sequence_lengths, numsteps, axis=2, scope=None)
	concatenate each two consecutive elements More...
def	nabu.neuralnetworks.components.ops.stack_seq (sequential, sequence_lengths, name=None)
	remove padding and stack sequences More...
def	nabu.neuralnetworks.components.ops.unstack_seq (nonseq, sequence_lengths, name=None)
	unstack sequences and add padding More...
def	nabu.neuralnetworks.components.ops.dense_sequence_to_sparse (sequences, sequence_lengths)
	convert sequence dense representations to sparse representations More...
def	nabu.neuralnetworks.components.ops.get_indices (sequence_length)
	get the indices corresponding to sequences (and not padding) More...
def	nabu.neuralnetworks.components.ops.pad_to (tensor, length, axis=0, name=None)
	pad the tensor to a certain length More...
def	nabu.neuralnetworks.components.ops.map_ta (fn, ta)
	apply fn to each element in tensorarray More...

Detailed Description

some operations

Function Documentation

dense_sequence_to_sparse()

def nabu.neuralnetworks.components.ops.dense_sequence_to_sparse	(		sequences,
			sequence_lengths
	)

convert sequence dense representations to sparse representations

Parameters

sequences	the dense sequences as a [batch_size x max_length] tensor
sequence_lengths	the sequence lengths as a [batch_size] vector

Returns

the sparse tensor representation of the sequences

get_indices()

def nabu.neuralnetworks.components.ops.get_indices

(

sequence_length

)

get the indices corresponding to sequences (and not padding)

Parameters

sequence_length

the sequence_lengths as a N-D tensor

Returns

A [sum(sequence_length) x N-1] Tensor containing the indices

map_ta()

def nabu.neuralnetworks.components.ops.map_ta	(		fn,
			ta
	)

apply fn to each element in tensorarray

Parameters

fn	the function to apply
ta	the tensorarray

Returns

the resulting tensorarray

pad_to()

def nabu.neuralnetworks.components.ops.pad_to	(		tensor,
			length,
			axis = 0,
			name = None
	)

pad the tensor to a certain length

• tensor: the tensor to pad
• length: the length to pad to, has to be larger than tensor.shape[axis]
• axis: the axis to pad
• name: the name of the operation

Returns

the padded tensor

pyramid_stack()

def nabu.neuralnetworks.components.ops.pyramid_stack	(		inputs,
			sequence_lengths,
			numsteps,
			axis = 2,
			scope = None
	)

concatenate each two consecutive elements

Parameters

inputs	A time minor tensor [batch_size, time, input_size]
sequence_lengths	the length of the input sequences
numsteps	number of time steps to concatenate
axis	the axis where the inputs should be stacked
scope	the current scope

Returns

inputs Concatenated inputs

           [batch_size, time/numsteps, input_size*numsteps]
@return        sequence_lengths    the lengths of the inputs sequences [batch_size]

stack_seq()

def nabu.neuralnetworks.components.ops.stack_seq	(		sequential,
			sequence_lengths,
			name = None
	)

remove padding and stack sequences

Parameters

sequential	the sequential data which is a [batch_size, max_length, dim]
tensor
sequence_lengths	a [batch_size] vector containing the sequence lengths
name	[optional] the name of the operation

Returns

non sequential data, which is a TxF tensor where T is the sum of all sequence lengths

unstack_seq()

def nabu.neuralnetworks.components.ops.unstack_seq	(		nonseq,
			sequence_lengths,
			name = None
	)

unstack sequences and add padding

Parameters

nonseq	the non sequential data which is a [sum(sequence_lengths) x dim] tensor
sequence_lengths	a [batch_size] vector containing the sequence lengths
name	[optional] the name of the operation

Returns

sequential data, which is a [batch_size, max_length, dim] tensor