Is tf.session thread-safe?

[jain....@gmail.com]

hi

I have come across a few mail threads discussing the same; but not in the exact context which I am looking at:

I tried to run 2 threads, both sharing the same session. One thread ingests data from (say a queue) and calls session.run() to train the model and the other thread runs inferencing using the same session. It seems to run in parallel, but:

1. I am not sure yet (need to verify) if there are thread safety issues (not sure what underlying operations/data structures are for updating model parameters) and if the result returned is correct or not.

2. If it is safe, is there a significant overhead wrt locking and slow down.

3. Are all the model weights updated using one single lock at the end of every training iteration.

Mridul

[Yaroslav Bulatov]

What you are describing under 1. is a very common scenario used during training, so it should work and be fairly optimized. There's no locking by default, which tends to work OK for gradient descent. Operations like "tf.assign_add" have `use_locking` attribute which you can set to ensure that variable doesn't change between reading and modification.

There's some more discussion on state of locking here: https://github.com/tensorflow/tensorflow/issues/6360

--
You received this message because you are subscribed to the Google Groups "Discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to discuss+unsubscribe@tensorflow.org.
To post to this group, send email to dis...@tensorflow.org.
To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/discuss/489ded71-051f-4307-a739-dab0493b2db4%40tensorflow.org.