search slow

[andyhahaha]

The paper said that NSGA-Net + macro search space takes 8 GPU-days NSGA-Net (6 @ 560) + cutout takes 4 GPU-days

But I run this searching code on Tesla P100-PCIE-16GB, the results are different. Macro takes 4 days Micro takes 41 days(30 mins for one network, total 0.5hr * 40 * 50 / 24hr = 41.6days)

This results is far away from the paper. Can you explain? Thanks!

[mikelzc1990]

The discrepancy in search time is due to the default hyper-parameters provided is not consistent with those used in the paper. Please use the following setting to run the search on micro space.

python search/evolution_search.py --search_space micro --init_channels 16 --layers 8 --epochs 20 --n_offspring 20 --n_gens 30

A bit more details, we halved the number of offspring created in each generation in micro search space case to reduce the search cost from 8 to 4 days. In the same time, we adjust the initial channels and number of epochs to match architecture's training time during search between macro search space models and micro search space models. But the archive size (population size) is the same between two cases.

[zkjst110]

when you run the macro search ,do you notice that the default parameter --n_nodes(number of nodes per phases) is four? the paper said

We set the number of phases np to three and the number of nodes in each phase no to six

so , i want to know if you have modified --n_nodes to six when you run the code for acquiring the time ( 8 GPU-days) .

thanks!