about batch size about kinetics-i3d HOT 4 OPEN

Yes, that is correct, but note that batch norm layers use just the data in each gpu so it's not completely equivalent to having a larger batch.

from kinetics-i3d.

@joaoluiscarreira if minibatch size is 6 for each gpu, it seems to be too small for training batch norm layer because batch norm is useful only when batch size is big enough.So can you explain it in detail?

from kinetics-i3d.

Note that all the layers are convolutional and batch norm accumulates statistics over the spatial and temporal dimensions so it's not as bad as if we would be talking about fully connected layers. I'm sure things would work out better if we could run batch norm over bigger batches but it's already ok.

from kinetics-i3d.

@joaoluiscarreira thanks for your help.

from kinetics-i3d.