Thinking about a better way to calculate computation latency of a client in a round

[yuxuan18]

Currently, FedScale calculates computation latency without considering the model complexity. This does not make sense if we have different models training in the same time or we want to compare the system latency of different models.

I propose to multiply the current formula by a ratio of the number of float point operations (FLOPs) of the used model and reference model.

[fanlai0990]

Agree to do so. But please do not change the augmentation_factor.

Instead, 1. remove these hard codes; 2. Multiply the compute_speed by the factor from FLOPs. Note that the current compute trace is from mobile_v2. So the factor should be FLOPs_of_new_model/FLOPs_of_mobile_v2