-1

I am using MultiprocessIterator from chainer to train an NN. The code is here.

The results are:

nproc = 1,  307.65s
nproc = 4,  318.85s
nproc = 8,  302.91s
nproc = 16, 318.18s

System info:

Ubuntu 14.04.5 LTS (GNU/Linux 4.4.0-104-generic x86_64)
CPU(s): 16
Model name: Intel(R) Xeon(R) CPU E5-2695 v4 @ 2.10GHz
VGA compatible controller: Cirrus Logic GD 5446

It seems MultiprocessIterator doesn't work. Anyone has a clue? Thank you in advance.

cmjdxy
  • 396
  • 1
  • 2
  • 15
  • Post the (relevant) code in your question please. – Julien May 08 '18 at 01:27
  • I tried, but the system said "It looks like your post is mostly code, please add some more details" and didn't approve my post. So I use a hyperlink. – cmjdxy May 08 '18 at 01:38

1 Answers1

0

Quote from: https://github.com/chainer/chainer/issues/4723

MultiiprocessIterator just (sometimes) speeds up picking up a batch from Dataset and it does not hasten ANY GPU computation. In most cases, the computational bottleneck is forwarding/backwarding, so it is only natural that MultiiprocessIterator has no effect on the total time.

kmaehashi
  • 879
  • 6
  • 10