Batching audio data in tensorflow

Question

I am trying to train some models on audio data. I wrote some code to load some mp3 files, split them up into short pieces (about 0.1 second each) and analyze these pieces in batches. So, I wrote this code.

import glob
import tensorflow as tf
from tensorflow.contrib import ffmpeg

def load(fname):
    binary = tf.read_file(fname)
    return ffmpeg.decode_audio(binary, file_format='mp3', samples_per_second=44100, channel_count=2)   

def preprocess(audio, seconds_per_sample=0.1, rate=44100):
    # pad to a with 1 second of silence front and back
    front = tf.zeros([rate, 2], dtype=audio.dtype)
    back = tf.zeros([rate - tf.mod(tf.shape(audio)[0], rate) + rate, 2], dtype=audio.dtype)
    audio = tf.concat([front, audio, back], 0)
    # normalize to 0 to 1 range
    audio = tf.add(audio, tf.abs(tf.reduce_min(audio)))
    audio = tf.multiply(audio, 1.0 / tf.reduce_max(audio))
    # [data, channels] => [samples, data, channels]
    audio = tf.reshape(audio, [-1, int(rate * seconds_per_sample), 2])
    return audio

tf.reset_default_graph()
with tf.Graph().as_default():
    # take files one by one and read data from them
    files = glob.glob('music/*.mp3')    
    queue = tf.train.string_input_producer(files, num_epochs=1)
    fname = queue.dequeue()
    audio = load(fname)
    audio = preprocess(audio)
    samples = tf.train.slice_input_producer([audio], num_epochs=1)
    batch = tf.train.batch(samples, 10)

    model = tf.identity(batch)

    init = [tf.global_variables_initializer(), tf.local_variables_initializer()]

    coord = tf.train.Coordinator()

    with tf.Session() as session:
        session.run(init)
        threads = tf.train.start_queue_runners(sess=session, coord=coord)
        for _ in range(10):
            try:
                result = session.run(model)
            except tf.errors.OutOfRangeError:
                coord.request_stop()
        coord.request_stop()
        coord.join(threads)

It seems pretty straight forward an similar approaches worked for me for my previous models. I reshape the audio data, so the first dimension becomes samples, use the slice input to queue samples up and then use batch() to feed the samples 10 at a time into the model. For simplicity, I left the model as an identity function. This code makes my python segfault somewhere deep inside tensorflow. Is there anything I am doing obviously wrong?

Here is the start of OSX crash report

Process:               Python [57865]
Path:                  /usr/local/Cellar/python3/3.6.1/Frameworks/Python.framework/Versions/3.6/Resources/Python.app/Contents/MacOS/Python
Identifier:            Python
Version:               3.6.1 (3.6.1)
Code Type:             X86-64 (Native)
Parent Process:        Python [57654]
Responsible:           Python [57865]
User ID:               502

Date/Time:             2017-04-12 16:07:13.318 -0400
OS Version:            Mac OS X 10.12.3 (16D32)
Report Version:        12
Anonymous UUID:        B5DE676B-FEC7-9626-B1CC-F392948D410C

Sleep/Wake UUID:       F3A5360E-B7A0-4675-9DC9-EAEE938E2E70

Time Awake Since Boot: 440000 seconds
Time Since Wake:       16000 seconds

System Integrity Protection: disabled

Crashed Thread:        16

Exception Type:        EXC_CRASH (SIGABRT)
Exception Codes:       0x0000000000000000, 0x0000000000000000
Exception Note:        EXC_CORPSE_NOTIFY

Application Specific Information:
abort() called

Edit: the issue I opened on GitHub was closed with no explanation, but "see issue tracker policy". I am not sure what else I can do here. If anyone has any light to shed on this problem, please do.

Answer 1

Before running the code You have posted on my computer, I had to add some MP3-file to "music" folder. I assume You have some audio there, but also, please pay attention to ffmpeg binary. Tensorflow asks for ffmpeg to be in /usr/local/sbin/ folder.

A Quick Workaround

A usual symbolic link worked for me.

ln -s /usr/bin/ffmpeg /usr/local/sbin/ffmpeg

If this answer is not helpful, then, please, provide more information by running the code in a terminal emulator and posting the traceback here.