FastBatch: A Multimodal Parallelized Neural Network Iterator for Python 2.7

Eugene Laksana, Sayan Ghosh, Stefan Scherer
University of Southern California, Institute for Creative Technologies

Description

A python-based neural network batch iterator with multimodal support. It works by first splitting the dataset into separate chunks and designating concurrent workers to populate a queue from their assigned chunks. A single process is responsible for retrieving batches from this queue via a get_batch function. Currently, this version has both numerical and lexical support.

Requirements

• Python 2.7x
• numpy
• h5py

Data Input Format

The iterator currently only supports the hdf5 file format. As an input parameter, FastBatch takes in the path to a file list which contains paths to the hdf5 files separated by newline characters. I have provided some functional examples of our input format under the ex folder.

Usage

There are currently two working versions of fastbatch: fastbatch and fastbatch_term.

fastbatch: continues to populate the concurrent queue indefinitely by looping over the file list.

fastbatch_term: returns a None tuple after the last element in the file list has been reached.

Both versions employ the same parameters:

Mandatory

file_list_name: path to the file list.

feat_dsname: the dataset name of the numerical features (ie. openSMILE_features.)

word_dsname: the dataset name of the lexical features

new_file_list_prefix: prefix for the file list copies that will be generated for each process to work on.

k: number of chunks to make. (will assign k workers to work on k chunks of the file list and populate the queue from their respective chunks.)

batch_size: number of datapoints per batch.

num_timesteps: number of batches desires at a time.

Optional

shift_step: number of shifts requested for the shifted array (default: 1)

word_to_id: Sayan's parameter. (default: None)

feat_list: This only works with our modified OpenSmile IS11 features. Leave at None if irrelevant (default: None)

remove_sp: removes sp, which are non-word sounds as labeled in the Fisher dataset. Leave false if irrelevant (default: False)

Return

orig_word_mat: original word matrix

shift_word_mat: shifted word matrix

feat_mat: feature matrix

hdf5_filepath_mat: hdf5 filepaths from which the original word matrix came from.

Sample Usage

from fastbatch import fastbatch
import time, sys, os

if __name__ == '__main__':
	file_list_name = './ex/file_list.txt'
	feat_dsname = 'openSMILE_features'
	word_dsname = 'words'
	new_file_list_prefix = './file_list_'
	k = 3
	batch_size = 20
	num_timesteps = 20
	shift_step = 1
	pdi = fastbatch(file_list_name, feat_dsname, word_dsname, new_file_list_prefix, k, batch_size, num_timesteps, shift_step, feat_list = ['f0', 'shimmer', 'jitter', 'voicing', 'rmsenergy'], remove_sp = True)
	time_count = 0
	st = time.time()
	for i in range(0, 100):
		sys.stdout.flush()
		a,b,c,d = pdi.get_batch()
		if i % 10 == 0:
			ed = time.time()
			print str(i) + 'th batch: ' + str(ed - st)
	ed = time.time()
	print ed - st
	sys.stdout.flush()