Skip to content

Cntk Speed Tests

Jump to bottom
BorjaFG edited this page Apr 2, 2019 · 3 revisions

On the following images, we have plotted the time it took in our experiments to train/evaluate a 3-layer network with different minibatch sizes. In all the cases, we used 10^6 tuples. Clearly, batching tuples results in much faster training/evaluation:

  • Training Training times

  • Evaluation Evaluation times

  • DQN update rule DQN update times

SimionSoft - Group of Computational Intelligence ( University of the Basque Country (UPV/EHU) )

Home

Clone this wiki locally