is the evalset_ and the step of evaluate(epoch); necessary if we dont care the loss rate for the online envitoment ?

[jackyhou2014]

is the evalset_ and the step of
evaluate(epoch);
necessary if we dont care the loss rate for the online enviroment ?

for without evalset, we can save a lot of time

[jackyhawk]

It seems that if u don't care the metrics (since we already care a lot of those all metrics at test time),
so we don't need this (evalSet_.empty() ? -1.0 : parallel_.mapReduce) ? just dont initialize the evalSet_ ?

[jackyhawk]

Any suggestions for that? thanks very much

[albietz]

Hi,

Typically the averaged test metrics are slower to compute, so this shouldn't be a bottleneck, especially when parallelized, but you can definitely comment it out if it slows things down for you. @chtran @JelleZijlstra can make the call on whether it's worth adding gflags for this.

-Alberto

[jackyhawk]

Thanks very much Alberto.

And the thing that I want to do is just remove the initializing the evalSet(not the test set) on the production environment ( as I don’t need to watch the metrics for the production results).

And as I review the code, it seemed that it would have no effect for the final result (removing the evalset). I am not clear whether that is the truth (since if i just remove the evalset and the metrics of the evalset, it would save about half of the whole time)

Thanks very much

[jackyhawk]

And it would be better if you provide gflags to choose whether use evalset or not,

Thanks veny much Alberto.

[jackyhawk]

And what's more, it would save a lot of memory