Hardware and training time?

[whatever60]

Great work.

But what's the hardware used to reproduce this? And how long did it take to train?

Thanks

[WangFeng18]

To perfectly reproduce the reported results, using 8xTesla V100 with total batch size of 1024 (or 16 x V100), costing about 40h. We have not try other settings.
However, I believe when using smaller batch size such as 512 or 256, you can still get reasonable results.

[whatever60]

Thank you.

That's indeed a fair amount of computation compared to traditional CNN.