0
$\begingroup$

It is often the case when the time vs memory trade-off is underestimated prior to using ML/DL for solving a particular task. Taking into account the type, size and format of the available data and also the available CPUs, GPUs and RAM as well, I wonder is there a general way to estimate prior to modelling:

1) How much computational time will be needed?

2) What will be the maximum amount of memory needed?

I assume a correct estimation of the aforementioned issues will be quite useful prior to modelling, taking into account also the time complexity of the algorithms to use as well.

Improve this question
$\endgroup$

1 Answer 1

Reset to default
0
$\begingroup$

From the world of Algorithms, we can borrow the "Complexity" estimation science to find out the required time (time complexity) and the memory (space complexity) to execute. Given most of the problems & algorithms in the ML world are from the NP family, it does not make sense to use complexity as is. Instead, we may use complexity estimation alongside heuristics to get close to the correct value of time and space complexities.

An empirical approach is to downsize your experiment to a small sample experiment and measure the wall time. When you plug in the parameters related to CPU, GPU speeds and memory consumed, you may guestimate the time and memory required for the scaled-up version of your experiment. This approach is gross approximation of the original problem, but a practical way to get the ballpark numbers.

Improve this answer
$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .