CODEWITHSUNDEEP
caffeconv-neural-network
SimpleNotGood
SimpleNotGood

Reputation: 345

Delay layer output for one iteration

I have a fcn which should analyse a video. I want to use the output of the network as input for the net in the next iteration. Is there a layer which can do this? If not, is it possible to write such a layer in a reasonable amount of time? If so, do you have any advices I should consider while porgramming the layer?

Upvotes: 0

Views: 24

Answers (1)

Shai
Shai

Reputation: 114786

It seems like you are looking for a "Recurrent" layer.
Note that having a recurrence in your net, the net now has a "cycle" in its underlying graph. Caffe usually uses "unrolling" to work around these cyclic graphs.

Upvotes: 2

Related Questions