Tried to write tfrecord w/ and w/o multithreading, and found the speed difference is not much (w/ 4 threads: 434 seconds; w/o multithread 590 seconds). Not sure if I used it correctly. Is there any better way to write tfrecord faster?
import tensorflow as tf
import numpy as np
import threading
import time
def generate_data(shape=[15,28,60,1]):
return np.random.uniform(size=shape)
def _bytes_feature(value):
return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))
def _int64_feature(value):
return tf.train.Feature(int64_list=tf.train.Int64List(value=[value]))
def write_instances_to_tfrecord(tfrecord_file, filenames):
tfrecord_writer = tf.python_io.TFRecordWriter(tfrecord_file)
for i, filename in enumerate(filenames):
curr_MFCC = generate_data()
curr_MFCC_raw = curr_MFCC.tostring()
curr_filename_raw = str(filename)+'-'+str(i)
example = tf.train.Example(features=tf.train.Features(
feature={
'MFCC': _bytes_feature(curr_MFCC_raw),
'filename': _bytes_feature(curr_filename_raw)
})
)
tfrecord_writer.write(example.SerializeToString())
tfrecord_writer.close()
def test():
threading_start = time.time()
coord = tf.train.Coordinator()
threads = []
for thread_index in xrange(4):
args = (str(thread_index), range(200000))
t = threading.Thread(target=write_instances_to_tfrecord, args=args)
t.start()
threads.append(t)
coord.join(threads)
print 'w/ threading takes', time.time()-threading_start
start = time.time()
write_instances_to_tfrecord('5', range(800000))
print 'w/o threading takes', time.time()-start
if __name__ == '__main__':
test()
When using python threading, due to the GIL restriction in the cPython implementation, the CPU utilization will be capped at 1 core. No matter how many threads you add, you will not see a speed up.
A simple solution in your case would be to use the multiprocessing module.
The code is almost exactly the same as what you have, just switch threads to processes:
from multiprocessing import Process
coord = tf.train.Coordinator()
processes = []
for thread_index in xrange(4):
args = (str(thread_index), range(200000))
p = Process(target=write_instances_to_tfrecord, args=args)
p.start()
processes.append(p)
coord.join(processes)
I tested this on my own tfrecord writer code, and got a linear scaling speedup. Total number of processes is limited by memory.
It's better to use Tensorflow computation graph to take advantage of multithreading since each session and graph can be run in different threads. With computation graph, it's about 40 times faster.
Related
I am using the following code to process some pictures for my ML project and I would like to parallelize it.
import multiprocessing as mp
import concurrent.futures
def track_ids(seq):
'''The func is so big I can not put it here'''
ood = {}
for i in seq:
# I load around 500 images and process them
ood[i] = some Value
return ood
seqs = []
for seq in range(1, 10):# len(seqs)+1):
seq = txt+str(seq)
seqs.append(seq)
# serial call of the function
track_ids(seq)
#parallel call of the function
with concurrent.futures.ProcessPoolExecutor(max_workers=mp.cpu_count()) as ex:
ood_id = ex.map(track_ids, seqs)
if I run the code serially it takes 3.0 minutes but for parallel with concurrent, it takes 3.5 minutes.
can someone please explain why is that? and present a way to solve the problem.
btw, I have 12 cores.
Thanks
Here's a brief example of how one might go about profiling multiprocessing code vs serial execution:
from multiprocessing import Pool
from cProfile import Profile
from pstats import Stats
import concurrent.futures
def track_ids(seq):
'''The func is so big I can not put it here'''
ood = {}
for i in seq:
# I load around 500 images and process them
ood[i] = some Value
return ood
def profile_seq():
p = Profile() #one and only profiler instance
p.enable()
seqs = []
for seq in range(1, 10):# len(seqs)+1):
seq = txt+str(seq)
seqs.append(seq)
# serial call of the function
track_ids(seq)
p.disable()
return Stats(p), seqs
def track_ids_pr(seq):
p = Profile() #profile the child tasks
p.enable()
retval = track_ids(seq)
p.disable()
return (Stats(p, stream="dummy"), retval)
def profile_parallel():
p = Profile() #profile stuff in the main process
p.enable()
with concurrent.futures.ProcessPoolExecutor(max_workers=mp.cpu_count()) as ex:
retvals = ex.map(track_ids_pr, seqs)
p.disable()
s = Stats(p)
out = []
for ret in retvals:
s.add(ret[0])
out.append(ret[1])
return s, out
if __name__ == "__main__":
stat, retval = profile_parallel()
stat.print_stats()
EDIT: Unfortunately I found out that pstat.Stats objects cannot be used normally with multiprocessing.Queue because it is not pickleable (which is needed for the operation of concurrent.futures). Evidently it normally will store a reference to a file for the purpose of writing statistics to that file, and if none is given, it will by default grab a reference to sys.stdout. We don't actually need that reference however until we actually want to print out the statistics, so we can just give it a temporary value to prevent the pickle error, and then restore an appropriate value later. The following example should be copy-paste-able and run just fine rather than the pseudocode-ish example above.
from multiprocessing import Queue, Process
from cProfile import Profile
from pstats import Stats
import sys
def isprime(x):
for d in range(2, int(x**.5)):
if x % d == 0:
return False
return True
def foo(retq):
p = Profile()
p.enable()
primes = []
max_n = 2**20
for n in range(3, max_n):
if isprime(n):
primes.append(n)
p.disable()
retq.put(Stats(p, stream="dummy")) #Dirty hack: set `stream` to something picklable then override later
if __name__ == "__main__":
q = Queue()
p1 = Process(target=foo, args=(q,))
p1.start()
p2 = Process(target=foo, args=(q,))
p2.start()
s1 = q.get()
s1.stream = sys.stdout #restore original file
s2 = q.get()
# s2.stream #if we are just adding this `Stats` object to another the `stream` just gets thrown away anyway.
s1.add(s2) #add up the stats from both child processes.
s1.print_stats() #s1.stream gets used here, but not before. If you provide a file to write to instead of sys.stdout, it will write to that file)
p1.join()
p2.join()
I am trying to use multiprocessing for the below code. The code seems to run a bit faster than the for loop inside the function.
How can I confirm I using the library and not the just the for loop?
from multiprocessing import Pool
from multiprocessing import cpu_count
import requests
import pandas as pd
data= pd.read_csv('~/Downloads/50kNAE000.txt.1' ,sep="\t", header=None)
data = data[0].str.strip("0 ")
lst = []
def request(x):
for i,v in x.items():
print(i)
file = requests.get(v)
lst.append(file.text)
#time.sleep(1)
if __name__ == "__main__":
pool = Pool(cpu_count())
results = pool.map(request(data))
pool.close() # 'TERM'
pool.join() # 'KILL'
Multiprocessing has overhead. It has to start the process and transfer function data via interprocess mechanism. Just running a single function in another process vs. running that same function normally is always going to be slower. The advantage is actually doing parallelism with significant work in the functions that makes the overhead minimal.
You can call multiprocessing.current_process().name to see the process name change.
I am following the principles laid down in this post to safely output the results which will eventually be written to a file. Unfortunately, the code only print 1 and 2, and not 3 to 6.
import os
import argparse
import pandas as pd
import multiprocessing
from multiprocessing import Process, Queue
from time import sleep
def feed(queue, parlist):
for par in parlist:
queue.put(par)
print("Queue size", queue.qsize())
def calc(queueIn, queueOut):
while True:
try:
par=queueIn.get(block=False)
res=doCalculation(par)
queueOut.put((res))
queueIn.task_done()
except:
break
def doCalculation(par):
return par
def write(queue):
while True:
try:
par=queue.get(block=False)
print("response:",par)
except:
break
if __name__ == "__main__":
nthreads = 2
workerQueue = Queue()
writerQueue = Queue()
considerperiod=[1,2,3,4,5,6]
feedProc = Process(target=feed, args=(workerQueue, considerperiod))
calcProc = [Process(target=calc, args=(workerQueue, writerQueue)) for i in range(nthreads)]
writProc = Process(target=write, args=(writerQueue,))
feedProc.start()
feedProc.join()
for p in calcProc:
p.start()
for p in calcProc:
p.join()
writProc.start()
writProc.join()
On running the code it prints,
$ python3 tst.py
Queue size 6
response: 1
response: 2
Also, is it possible to ensure that the write function always outputs 1,2,3,4,5,6 i.e. in the same order in which the data is fed into the feed queue?
The error is somehow with the task_done() call. If you remove that one, then it works, don't ask me why (IMO that's a bug). But the way it works then is that the queueIn.get(block=False) call throws an exception because the queue is empty. This might be just enough for your use case, a better way though would be to use sentinels (as suggested in the multiprocessing docs, see last example). Here's a little rewrite so your program uses sentinels:
import os
import argparse
import multiprocessing
from multiprocessing import Process, Queue
from time import sleep
def feed(queue, parlist, nthreads):
for par in parlist:
queue.put(par)
for i in range(nthreads):
queue.put(None)
print("Queue size", queue.qsize())
def calc(queueIn, queueOut):
while True:
par=queueIn.get()
if par is None:
break
res=doCalculation(par)
queueOut.put((res))
def doCalculation(par):
return par
def write(queue):
while not queue.empty():
par=queue.get()
print("response:",par)
if __name__ == "__main__":
nthreads = 2
workerQueue = Queue()
writerQueue = Queue()
considerperiod=[1,2,3,4,5,6]
feedProc = Process(target=feed, args=(workerQueue, considerperiod, nthreads))
calcProc = [Process(target=calc, args=(workerQueue, writerQueue)) for i in range(nthreads)]
writProc = Process(target=write, args=(writerQueue,))
feedProc.start()
feedProc.join()
for p in calcProc:
p.start()
for p in calcProc:
p.join()
writProc.start()
writProc.join()
A few things to note:
the sentinel is putting a None into the queue. Note that you need one sentinel for every worker process.
for the write function you don't need to do the sentinel handling as there's only one process and you don't need to handle concurrency (if you would do the empty() and then get() thingie in your calc function you would run into a problem if e.g. there's only one item left in the queue and both workers check empty() at the same time and then both want to do get() and then one of them is locked forever)
you don't need to put feed and write into processes, just put them into your main function as you don't want to run it in parallel anyway.
how can I have the same order in output as in input? [...] I guess multiprocessing.map can do this
Yes map keeps the order. Rewriting your program into something simpler (as you don't need the workerQueue and writerQueue and adding random sleeps to prove that the output is still in order:
from multiprocessing import Pool
import time
import random
def calc(val):
time.sleep(random.random())
return val
if __name__ == "__main__":
considerperiod=[1,2,3,4,5,6]
with Pool(processes=2) as pool:
print(pool.map(calc, considerperiod))
I have a function that yields lines from a huge CSV file lazily:
def get_next_line():
with open(sample_csv,'r') as f:
for line in f:
yield line
def do_long_operation(row):
print('Do some operation that takes a long time')
I need to use threads such that each record I get from the above function I can call do_long_operation.
Most places on Internet have examples like this, and I am not very sure if I am on the right path.
import threading
thread_list = []
for i in range(8):
t = threading.Thread(target=do_long_operation, args=(get_next_row from get_next_line))
thread_list.append(t)
for thread in thread_list:
thread.start()
for thread in thread_list:
thread.join()
My questions are:
How do I start only a finite number of threads, say 8?
How do I make sure that each of the threads will get a row from get_next_line?
You could use a thread pool from multiprocessing and map your tasks to a pool of workers:
from multiprocessing.pool import ThreadPool as Pool
# from multiprocessing import Pool
from random import randint
from time import sleep
def process_line(l):
print l, "started"
sleep(randint(0, 3))
print l, "done"
def get_next_line():
with open("sample.csv", 'r') as f:
for line in f:
yield line
f = get_next_line()
t = Pool(processes=8)
for i in f:
t.map(process_line, (i,))
t.close()
t.join()
This will create eight workers and submit your lines to them, one by one. As soon as a process is "free", it will be allocated a new task.
There is a commented out import statement, too. If you comment out the ThreadPool and import Pool from multiprocessing instead, you will get subprocesses instead of threads, which may be more efficient in your case.
Using a Pool/ThreadPool from multiprocessing to map tasks to a pool of workers and a Queue to control how many tasks are held in memory (so we don't read too far ahead into the huge CSV file if worker processes are slow):
from multiprocessing.pool import ThreadPool as Pool
# from multiprocessing import Pool
from random import randint
import time, os
from multiprocessing import Queue
def process_line(l):
print("{} started".format(l))
time.sleep(randint(0, 3))
print("{} done".format(l))
def get_next_line():
with open(sample_csv, 'r') as f:
for line in f:
yield line
# use for testing
# def get_next_line():
# for i in range(100):
# print('yielding {}'.format(i))
# yield i
def worker_main(queue):
print("{} working".format(os.getpid()))
while True:
# Get item from queue, block until one is available
item = queue.get(True)
if item == None:
# Shutdown this worker and requeue the item so other workers can shutdown as well
queue.put(None)
break
else:
# Process item
process_line(item)
print("{} done working".format(os.getpid()))
f = get_next_line()
# Use a multiprocessing queue with maxsize
q = Queue(maxsize=5)
# Start workers to process queue items
t = Pool(processes=8, initializer=worker_main, initargs=(q,))
# Enqueue items. This blocks if the queue is full.
for l in f:
q.put(l)
# Enqueue the shutdown message (i.e. None)
q.put(None)
# We need to first close the pool before joining
t.close()
t.join()
Hannu's answer is not the best method.
I ran the code on a 100M rows CSV file. It took me forever to perform the operation.
However, prior to reading his answer, I had written the following code:
def call_processing_rows_pickably(row):
process_row(row)
import csv
from multiprocessing import Pool
import time
import datetime
def process_row(row):
row_to_be_printed = str(row)+str("hola!")
print(row_to_be_printed)
class process_csv():
def __init__(self, file_name):
self.file_name = file_name
def get_row_count(self):
with open(self.file_name) as f:
for i, l in enumerate(f):
pass
self.row_count = i
def select_chunk_size(self):
if(self.row_count>10000000):
self.chunk_size = 100000
return
if(self.row_count>5000000):
self.chunk_size = 50000
return
self.chunk_size = 10000
return
def process_rows(self):
list_de_rows = []
count = 0
with open(self.file_name, 'rb') as file:
reader = csv.reader(file)
for row in reader:
print(count+1)
list_de_rows.append(row)
if(len(list_de_rows) == self.chunk_size):
p.map(call_processing_rows_pickably, list_de_rows)
del list_de_rows[:]
def start_process(self):
self.get_row_count()
self.select_chunk_size()
self.process_rows()
initial = datetime.datetime.now()
p = Pool(4)
ob = process_csv("100M_primes.csv")
ob.start_process()
final = datetime.datetime.now()
print(final-initial)
This took 22 minutes. Obviously, I need to have more improvements. For example, the Fred library in R takes 10 minutes maximum to do this task.
The difference is: I am creating a chunk of 100k rows first, and then I pass it to a function which is mapped by threadpool(here, 4 threads).
I am learning about QRunnable and I have the following code:
from PyQt5.QtCore import QThreadPool, QRunnable
class SomeObjectToDoComplicatedStuff(QRunnable):
def __init__(self, name):
QRunnable.__init__(self)
self.name = name
def run(self):
print('running', self.name)
a = 10
b = 30
c = 0
for i in range(5000000):
c += a**b
print('done', self.name)
pool = QThreadPool.globalInstance()
pool.setMaxThreadCount(10)
batch_size = 100
workers = [None] * batch_size
for i in range(batch_size):
worker = SomeObjectToDoComplicatedStuff('object ' + str(i))
workers[i] = worker
pool.start(worker)
print('All cued')
pool.waitForDone()
# processing the results back
for i in range(batch_size):
print(workers[i].name, ' - examining again.')
I see that indeed there are different processes being alternated, but all is happening on a single core.
How can I make this code run using all the processor cores?
PS: This code is just a simplification of a super complicated number crunching application I am making. In it, I want to to do Monte Carlo in several threads and the worker itself is a complex optimization problem.
I have tried the python multiprocessing module but it doesn't handle scipy too well.
Not sure how much use this will be, but a multiprocessing version of your example script would be something like this:
from multiprocessing import Pool
class Worker(object):
def __init__(self, name):
self.name = name
def run(self):
print('running', self.name)
a = 10
b = 30
c = 0
for i in range(5000000):
c += a**b
print('done', self.name)
return self.name, c
def caller(worker):
return worker.run()
def run():
pool = Pool()
batch_size = 10
workers = (Worker('object%d' % i) for i in range(batch_size))
result = pool.map(caller, workers)
for item in result:
print('%s = %s' % item)
if __name__ == '__main__':
run()
How can I make this code run using all the processor cores?
Using PyQt (QRunner/QThread and likely), I think it's almost impossible because they (the python version, not the C++) are using the GIL.
The easiest solution would be to use multiprocessing, but since you have some problem using it along scipy you should look for some non-standard library.
I suggest you to take a look at ipyparallel, AFAIK they're developed under the same umbrella, so they're likely to work seamlessy.