Difficulty understanding FLOPS in this scenario

Question

Given FLOPS are the floating point operations per second, would that not be dependent on the power of the machine rather than the model and how many parameters it has? What am I missing here? Screenshot is from "EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks". Thanks.

Maxim Lyuzin · Accepted Answer · 2022-06-04 17:59:36Z

1

Some hardware manufacturers specify FLOPS as the performance metric. On the other hand, you can calculate approximate FLOPs value for your model, for regular (not depthwise) convolutional layer it would be (according to this):

where 2 is for two different types of instructions: multiplying and accumulating.

You need to keep in mind that low FLOPS value do not automatically provide high performance.

edited Jun 4, 2022 at 17:59

answered Jun 3, 2022 at 13:42

Maxim Lyuzin

3973 silver badges8 bronze badges

I see, so should that be FLOPs (total operations) in the final column rather than FLOPS (total operations per second). And would it then mean the FLOPs for a single pass of the CNN? Thanks
– numunu
Commented Jun 4, 2022 at 2:10
@numunu Yep you are right, FLOPS for hardware and FLOPs for neural net forward pass. Edited that in the answer
– Maxim Lyuzin
Commented Jun 4, 2022 at 18:00

Add a comment |

Collectives™ on Stack Overflow

Difficulty understanding FLOPS in this scenario

1 Answer 1

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

Your Answer

Sign up or log in

Post as a guest

Related