test iterations

[tamar elazari]

Hi all!

I can't find what is the meaning behind the iterations for testing a net.
In the mnist example it is clear that the accuracy and loss are changing when iterations number is changing, but not when repeating a run for the same number of iterations. that rules out randomization...

Can anyone explain this parameter for me?

Thanks.

[Ran Manor]

The number of test iterations means how many time to run the test network when testing data.

So if you have batch_size=10 and test_iterations=500, at each test_interval you will run 5000 samples in the test network and compute the accuracy on their output.

[tamar elazari]

I'm not sure I understand - I have already trained my network, now I would like to test it. If I have batch_size=10, then I take 10 data items at each run, right? why do I need another parameter? since it is called iterations, my original thought was that it has to do with some algorithmic improvement of the results. does it? but what you wrote raises another question - is it just a way to tell the network how many sample to test on?