Closed hfarhidzadeh closed 6 years ago
That's a good question. Actually it is crucial to have a decent number of channels to achieve performance. What I suggest you to do is to keep the first layer at least 16 and then try to vary the number of channels in the other layers systematically and perform the experiments to see which combination works best for you.
Number of channels (kernels) is a hyper parameter of the network and it needs to be set by experimenting in cross-validation (or to be picked reasonably without going through a rigorous testing).
Fausto Milletari
On Jun 5, 2018, at 6:48 AM, hfarhidzadeh notifications@github.com wrote:
Hi,
How can I cut the channel size of each level to half and preserve the performance of network? what I meant is turning 16, 32, 64, 128 and 256 (in deepest layer) to 8, 16, 32, 64 and 128.
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/faustomilletari/VNet/issues/50, or mute the thread https://github.com/notifications/unsubscribe-auth/AMtsvlvNZseICKXY2o6C4viudBh73IzBks5t5owQgaJpZM4Ua3Db.
Hi,
How can I cut the channel size of each level to half and preserve the performance of network? what I meant is turning 16, 32, 64, 128 and 256 (in deepest layer) to 8, 16, 32, 64 and 128.