TDNN

[aguha65]

Is it true that there is no support for time-delayed neural nets in Caffe? I am interested in training a net with variable size input (I cannot scale them all to the same size) where I can set the targets dynamically based on all the output activations.

[longjon]

For variable size inputs, see #594 (coming soon). (Or, depending on your application, you may find that simple padding suffices.)

[aguha65]

That's great! I saw #594, but wasn't sure if it will allow TDNN. Just to make sure we are on the same page, this is what I want to do: the input would be a 2D array M X N. I will have several hidden layers (some convolutional). The output would be m X n. Then I will be setting the m X n targets based on the entire output array and then backward feed.

[shelhamer]

Ask modeling questions on the caffe-users mailing list please!

For the loss over the whole sequence you will have to develop that yourself. Otherwise Caffe does recurrent models through weight sharing, defining a model for each length, or making use of variable size inputs in #594.