Can we disable tensorflow's theading?

[ghost]

I'm running tensorflow in a simulator that does not support mult-threading. Is it possible to disable tensorflow's multi-threading?

[alextp]

No. It is possible to use a single executor thread and a single eigen thread, but the API as designed is likely to deadlock if you have no threads.

[ghost]

Can you elaborate on how to use a single executor thread and a single eigen thread? I tracked down the error and the deepest failure point i can find is "impl_.reset(new ThreadPool::Impl(env, thread_options, "tf_" + name, num_threads, low_latency_hint));"

[alextp]

The ConfigProto you pass to the first session you build in a process lets you specify how many dispatch and compute threads to use.

[ghost]

I forgot to mention I'm using C++., Do you know how to do that in C++? Also to confirm, you mean it is impossible to have no thread, at least two treads, one executor and one eigen are needed, right?

[alextp]

In C++ you can also use a configproto in sessionoptions: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/public/session_options.h

The minimum I know you can have is 3 threads: the two you mentioned and the one which is calling tf code.

[ghost]

I changed num_threads at line 116 in threadpool.cc to 0 and I bypassed the threading error but of course when I do sesson.run() I got pointer pointing to invalid address.

[alextp]

0 is not supported, as I said above. It'll only work if each number of threads is at least 1.

…

On Tue, Jun 27, 2017 at 7:38 AM, kyo744 ***@***.***> wrote: I changed num_threads at line 116 in threadpool.cc to 0 and I bypassed the threading error but of course when I do sesson.run() I got pointer pointing invalid address. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#11066 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAATxd2pat4f0M-Kw1QTvGnssMqzz2Z1ks5sIRP2gaJpZM4OFtTa> .

-- - Alex

[mrlzla]

try this:

tensorflow::SessionOptions GetSessionOptions()
{
	tensorflow::SessionOptions sessionOptions;
	tensorflow::ConfigProto& config = sessionOptions.config;
	config.set_intra_op_parallelism_threads(1);
	config.set_inter_op_parallelism_threads(1);
	return sessionOptions;
}

std::unique_ptr<tensorflow::Session> session;
session.reset(tensorflow::NewSession(GetSessionOptions()));

[ghost]

I've tried that. Any non zero value would cause my simulator to crash.

[ali01]

This question is better asked on StackOverflow since it is not a bug or feature request. There is also a larger community that reads questions there. Thanks!