Tensorflow input pipeline for distributed training

Question

I am trying to figure out how to setup my input pipeline for tensorflow in distributed training. It's not clear whether the readers will read from a single process and send the data to all workers or each server will start it's own input pipeline? How do we ensure that every worker has a different input going to it?

Answer 1

I will give an example of how I do it:

import tensorflow as tf
batch_size = 50
task_index = 2
num_workers = 10
input_pattern = "gs://backet/dir/part-00*"

get all names of files in the bucket that correspond to input_pattern

files_names = tf.train.match_filenames_once(
                input_pattern, name = "myFiles")

select names for worker task_index . tf.strided_slice is like slice for lists: a[::,task_index] (select every task_index th file for worker task_index )

to_process = tf.strided_slice(files_names, [task_index],
                 [999999999], strides=[num_workers])
filename_queue = tf.train.string_input_producer(to_process,
                     shuffle=True, #shufle files
                     num_epochs=num_epochs)

reader = tf.TextLineReader()
_ , value = reader.read(filename_queue)
col1,col2 = tf.decode_csv(value,
        record_defaults=[[1],[1]], field_delim="\t")

train_inputs, train_labels = tf.train.shuffle_batch([col1,[col2]],
        batch_size=batch_size,
        capacity=50*batch_size,
        num_threads=10,
        min_after_dequeue = 10*batch_size,
        allow_smaller_final_batch = True)

loss = f(...,train_inputs, train_labels)
optimizer = ...

with tf.train.MonitoredTrainingSession(...) as mon_sess:
    coord = tf.train.Coordinator()
    with coord.stop_on_exception():
        _ = tf.train.start_queue_runners(sess = mon_sess, coord=coord)
        while not coord.should_stop() and not mon_sess.should_stop():
            optimizer.run()

I'm not sure my method is the best way to implement input pipeline in case of distributed TensorFlow implementation because each worker reads names of all files in the bucket

---

Good lecture about input pipeline in TensorFlow: http://web.stanford.edu/class/cs20si/lectures/notes_09.pdf