distributed-learning-contributivity icon indicating copy to clipboard operation
distributed-learning-contributivity copied to clipboard

Published 20 hours ago • verified publisher icon LabeliaLabs

Profile timing for minibatch federated learning

Open RomainGoussault opened this issue 4 years ago • 4 comments

Increasing the number of minibatch seems to increase the time to do a federated run. This might be due to the process of:

  • creating the model
  • setting the weights
  • compiling the model

image

RomainGoussault avatar Apr 03 '20 15:04 RomainGoussault

This can help: https://blog.tensorflow.org/2020/04/introducing-new-tensorflow-profiler.html?m=1

RomainGoussault avatar Apr 16 '20 15:04 RomainGoussault

image _make_train_function should be investigated

RomainGoussault avatar Apr 23 '20 07:04 RomainGoussault

@RomainGoussault do you think it would be interesting to profile again as you did above?

bowni avatar Jun 12 '20 12:06 bowni

Yes it would be interesting to use the tensorflow profiler : https://blog.tensorflow.org/2020/04/introducing-new-tensorflow-profiler.html?m=1

RomainGoussault avatar Jun 12 '20 17:06 RomainGoussault